From Memory Formation to Governance: A Panorama of Agent Memory

Long-term memory is not simply archived, but is closer to the engineering portal for continuous learning.

For a long time, Age Memoory was always given an equal value to an external database problem. This captures the locals, but it misses the focus. Memoory is never just a reservoir, but a circular mechanism that allows long-term cognitive formation, activation, and subsequent renewal.

If you look at the connection to the article I wrote before, it'll be clearer. Yes.From the cognitive structure of an intelligent body to the framework of an intelligent bodyAnd I'm using Coala to discuss the memory level of the smart body -- Memoory -- which is the key question that comes out of it: It is both part of the cognitive architecture and a long-term engineering system.

This article seeks to make up a relatively complete picture of the current, more fragmented discussion of Argentina Memoory. The text is presented in four steps: first, to clarify several common misconceptions — in particular, to tear memories down to the three layers of context, external memory, parameter memory, and to describe which layer the body of the document is located; then to describe the anatomical structure of the memory system (what type of memory, how it is activated and written); then to review how this map was built step by step; and then to give an assessment — and then to explain why today's assessment is still not up to the front of the Memoory and to try to design a systematic blueprint for evaluation.

Here, the border is drawn: the paper discusses how long-term memory is formed, organized, updated, disabled and evaluated; and Working Memory is the operational entry point for long-term memory impact actions, not the object to be developed. The current reasoning of how to select, compress, isolate and dispatch information is another independent line, and I put it in the middle of a new line."Context is All You Need: Context Project for Smart Bodies"This is a special development, not to mention here.

If one is to sum up in one sentence, the competition point for Agent Memoory is no longer a memory bank, but rather whether it can be kept, correct, and manageable in the long term.

Memoory, not a database, but a cognitive loop.

First, it is concluded that the focus of Memoory is not “how much remains”, but whether long-term memory can create a stable entry, movement, exit mechanism: what is left behind, when is activated, and how the operation is then amended and governed.

Why doesn't it equal RAG?

RAG certainly matters, but it solves only the search side of the problem--"What knowledge is needed now?", not"What's worth keeping for a long time?""How old memories are corrected""What's out of date?""Experience should be written as a rule or as a fact."I'm sorry. By equating memory directly to RAG, it will simplify the system to one-time reading. A Memoory system also contains, at a minimum, a compression of current processes of dialogue and action, a deposition of long experience, an update of Persona, the decomposition of conflict information, and the oblivion of ineffective content.

Why do long-term memories have to pass through work Memoory to be effective?

Long-term memory does not directly involve action. Whether it is stored in vectors, maps, databases or files, it must be ultimately retrieved, screened, compressed and then entered into the context of the current decision-making cycle. This is the location of the Working Memoory. It's like a limited-capacity workstation where the Persona constraints, the context of the current conversation, external feedback, knowledge and experience recalled from long-term memory all meet. The model is not about all memories, but about the information that is allowed to be present at this time on this desk. The effectiveness of Memoory is ultimately reflected in whether long-term memory can be activated at the right time and enters Working Memoory in a manner that is sufficiently restrained.

Three layer memory and cost structure

Before entering the subject, there is one of the most confusing divisions to be made clear: when we say"Here, put in a memory for Agent."Where exactly does knowledge live? By"Location of knowledge repository + Cost structure for writing and calling"The memory can be divided into three layers, complementary rather than alternative, depending on the scene.

L1 Full Context: Knowledge lives directly in the current context window, representing technology with long context and KV-cache / prompt Caching. It's cost-structured.Zero training, deduction period token-paying on a continuous basis, and degradation as history grows(context rot). It's essentially the physical base of Working Memoory -- it determines how much the workstation can be down and how long it can be. This layer is part of the context-based movement of running time and is not the object of this work.
L2 External Memory / Non Parameter Memory (RAG-like): Knowledge is stored in visible form outside the model - documents in vector banks, rules in file systems, physical relationships in knowledge maps. It can be written instantaneously, accurately updated, searchable on demand and fully accessible to humans.This floor is the main character of this paper.。
L3 Parameter Memory (Paramec Memoory): Knowledge is codified into model weights, and the ability to adapt by training to compress repeat experiences into models is made effective without a need for a visible search. It is extremely low-cost and responsive, but very costly to write (needs minor or ongoing training), has a long updating cycle and is not transparent and auditable.

Three layers in parallel, the relationship between parameter memory and external memory is essentially one.The trade-off between the cost of reasoning and the cost of trainingL3 moves costs forward to the training stage in exchange for immediate call at almost zero cost for the reasoning period; L2 leaves costs in the reasoning period (retribution + context occupation) in exchange for immediate and controllable writing and updating; L1 completely abandons incremental loading and avoids cost loss through better context technology and caches. The hierarchy therefore depends entirely on the scene — the high frequency, the stable, the knowledge that is used on a large scale — which tends to sink into L3, and the knowledge that is needed to be written, accurately updated, auditable and personalized, which tends to remain in L2. Probably just once left in L1. The parameter memory nodes that were drawn in the cognitive cycle map ahead of us were L3: it was a hidden influence, but not an inclusive route.

It's L3 worth a few words.Training routesBecause there's a common misconception: since you want to make models,"Remember."A new set of knowledge, just take these documents for continuous pre-training. In fact, this path is poor — while the small-scale language often shows up only once or twice each fact, and the large-scale pre-training allows models to absorb a fact that is repeated in multiple terms; direct renewal is inefficient and easily catastrophic. The more feasible route isSynthesizing a set of targeted training data from source files/problems before doing it。Synthetic Continued Pretraining(EntiGraph) is the representative of this idea: it draws significant entities from small source languages and continues to train them by generating a larger, more diverse and more learningable synthetic between entities; post-training models do not need to answer the questions in the original language, and when the original text is available for reasoning, the parameterized knowledge is also relevant to RAG Double-hicking supersI'm sorry. As for the specific parameteration of the means of writing,LoRA and Prefix-Tuning I'm on the mechanism.PEFT: Efficient fine-tuning of parametersIt's been developed, and only the special value of it in the context of the Memoory language is shown: today, which is now popular in KV-cache, Prefix-Tuning trained to match the prefix with the prefix that is used for cache recovery with natural alignment, so that a fraction of the parameter memory can be mounted at low cost; LoRA provides a sort of plugged-up, supermutable, low-strength increase that matches"Switching the memory of different parameters by scene"The need.

But it's important to stress that the arguments are equally inconvenient.TimescaleThis is a process of memory correction and reversal over time, and the rewriting of parameterized knowledge, which means that the synthesis of training data will include time validity — a difficult task, which is far from being resolved at this time. In fact,Time is a three-tiered common problem.L1 is shown as invalidation of the old context and context rot, L2 as a conflict between old memory and new facts, and L3 as a copy of the memory editing of the weight-level is extremely difficult to precise. The post will be re-examined.

The real hard part is reading and writing and governance.

Once the long-term memory is put back into the life cycle, the priority of the discussion is reordered. The structure of the diagram is more dependent on the quality of the writing; the search algorithm is more dependent on the quality of the writing."When to write, when to change, when to delete."It's also to be seen. The problem of memoory is never just a question of storage structure, but rather a matter of life cycle management — it requires the system to deal with activation, writing, updating, deleting, and then also with problem, conflict resolution, time effectiveness and rollback.

After understanding this premise, we are going to go into the concrete roll-out of the landscape.

A panorama of Argentina Memoory

Memoory is a cognitive cycle, not data. Library

Memoory is not a memory retrieval and database, but a complete cognitive process that allows long-term memory to grow and modify in action. Agent retrieves the knowledge and experience needed from long-term memory and loads it into the Working Memoory context -- This is...Enter.; Agent reasoned and acted in this context, generating new experiences and facts; then refined these new information, rewritten it back to the long-term memory, and updated Persona - This is...Leave.I'm sorry. The entrance memory activation, exit writing and updating, and Working Memoory is the running-time entry for this cognitive closed ring, not a certain impervious cache layer.

flowchart LR
    LTM["长期记忆<br/>──────<br/>Persona · 对话 · 经验 · 知识<br/>载体：图 / DB / 文件系统"]
    READ["Memory Activation / Agentic 检索<br/>──────<br/>结构化索引 (文件系统·LSP)<br/>语义索引 (向量检索)<br/>字符匹配 (Grep)"]
    WM["Working Memory<br/>──────<br/>Persona 约束边界<br/>检索知识 + 对话上下文<br/>Compact 压缩维护容量"]
    ACT["Agent 行动<br/>──────<br/>推理 · 对话 · 工具调用<br/>产生新经验与事实"]
    WRITE["Memory Formation / Update<br/>──────<br/>提炼要点 · 与已有记忆去重<br/>选择写入结构<br/>更新 Persona · 沉淀经验"]

LTM --&gt;|进场| READ --&gt; WM
WM --&gt; ACT --&gt;|退场| WRITE --&gt; LTM

PM[&quot;参数记忆&lt;br/&gt;──────&lt;br/&gt;模型权重&lt;br/&gt;成本高 · 非普惠&quot;] -.-&gt;|隐式| ACT

style WM fill:#f59e0b,color:#000,stroke:#b45309,stroke-width:4px
style LTM fill:#6366f1,color:#fff,stroke:#4f46e5
style READ fill:#10b981,color:#fff,stroke:#059669
style ACT fill:#64748b,color:#fff,stroke:#475569
style WRITE fill:#10b981,color:#fff,stroke:#059669
style PM fill:#cbd5e1,color:#334155,stroke:#94a3b8,stroke-dasharray:5 5

Working Memory: is the hub of the entire system. All memories are retrieved here, all new experiences are refined and returned, Compact maintains its limited capacity
Long-term memory: long-term memory type stored in graphic/DB/file system
Agenic Retrieval/Rewriting: make up both reading and writing ends. Retrieval layer is managed by Agent, which is the three types of indexing technique; writing layers determine what is worth remembering and what structure to store
Parameter Memory: Inhibitive influence behaviour, high cost, non-parametric memory is the inclusive route

And here's the functioning skeleton of the Memoory system -- the cognitive cycle defines the flow of entry, movement, exit. The following is the filling of the skeleton: what kind of memory can Agent be broken down into, and what functions they each perform?

Five memory types

When we ask,"What part of Agent's memory should be broken down?"At that time, current practice points to five levels:Persona Memory, Working Memoory, Scenario Memory, Experience Memory and Semantic KnowledgeI'm sorry. These five are not parallel lists, but each has different cognitive functions: Persona defines the behavioral boundary - who and by what attitude is Agent; Working Memoory is the operational entry point for current decision-making; dialogue memory is the lasting chemical product of the compacted workstation, preserving the interactive link; empirical memory is sinking"What do we do next time?"and the human rights of the child; semantics"About what the world knows?"I'm sorry. Of these, Working Memoory is closer to the operational state interface, and Persona, scenario memory, empirical memory and semantic knowledge form the backbone of long-term memory management.

Persona Memory: is the foundation of the personality of Agent. It defines who Agent is — including the initial norms of personality creation, response and reflection, and the preferences and styles that evolve over time in a long interaction. Persona is just needed for scenarios that require a role-player, such as Openclaw, Nanobot, World Simulation; even the tool-type Agent, it will sink the hidden Persona - user preferences in long-term service feedback - as part of the Agent behaviour code itself. Persona remembers the closest thing in the CoALA framework to the concept of program memory, but it is not entirely accommodated by this framework: program memory emphasizes how to do it, and Persona also sets out what to do and what not to do. I have chosen to list it here separately, because it has a separate value for maintenance.

Working Memory: is the current cognitive desk of Agent. It is not a long-term storage, but a dynamic area of limited capacity — all the knowledge retrieved from long-term memory, the context of the current dialogue, and the constraints imposed by Persona — all gathered here to form the input of Agent's every step of reasoning. The continuity of dialogue depends heavily on the management of Work Memoory: the limited capacity of the information is maintained without loss through compressed technologies such as Compact, enabling Agent to remain coherent in long conversations. Early memory research is almost equivalent to context management for dialogue, essentially addressing the capacity bottlenecks of Working Memoory; but in this paper it is mainly used as an entry point for long-term memory to take effect, rather than as a separate object for work.

Episodic Memory: is a permanent chemical product of the long-term memory that is compressed by Work Memoory, and is a memory in time series form. Working Memoory is a working-state workstation, which exists only at the current session; and scenario memory survives through the session, as part of a long-term memory. In a long conversation, Agent and the users will have a large amount of information exchange, some of which will be refined to include experience or semantic knowledge, and some will update Persona, but there will always be a considerable amount of content that does not fall into any of the above categories - they are the context of the discussions, the ad hoc intent of the users, the undetermined direction of the exploration, or simply, the direction of the discussions."We talked about it."The fact itself. These elements may still be of value to users, but they do not deserve to be formally archived as experience or knowledge. When Working Memory had to compress the historical context because of capacity constraints, these elements, as products of compression and refining, were permanently translated into a memory of dialogue, allowing Agent to continue the previous conversation in the follow-up session. It should be engineeringly separated from semantic knowledge, which is a fact of the world, of the enduring validity of the inter-conference, and scenario memory, of the context that interacts with itself and dies with the life cycle of the session. The initial phase of the Memoory study, which was almost equivalent to the continuous study of dialogue, addressed the very same level of problem.

Experience Memory: It's the operational knowledge that Agent learned from his actions. When Agent is no longer limited to chatting but actually enters the production environment to perform its tasks, it accumulates successes and failures in the course of its operations - they can automatically extract sedimentation (e.g., automatic evolutionary mechanisms) or be manually filled by humans (e.g., Skylls and Rules in Claude Code). The empirical memory is the key increment that Agent has evolved from a dialogue tool to an operational entity, and it answers the question about what to do in a similar setting.

Semantic Memory: Agent, the world's factual perception - the relationship between entities, the conceptual system in the field, the specific facts that users have informed. It is different from the operational orientation of experience memory, but it is a pure knowledge of what: know the address of the user, know the call agreement of an API, know the causal relationship between the two concepts. Semantic knowledge is one of the main objects of retrieval and the type of memory that can be most easily organized in structured ways (knowledge mapping, databases).

Storage Structure Selection

The three-layer framework that preceded has defined the boundary between the memory of parameters and the memory of non-parametrics, and also indicates that the subject matter of this paper is located in the layer of the memory of non-parametrics. What is further asked here is what structure should be used to organize the non-parametric memory inside?

Memory requires structures, possibly maps, structured databases or document systems — but structures are not ends in themselves. ** More important than the structure, how it grows, how it is modified, how it is retrieved. ** A carefully designed knowledge map, without an effective updating mechanism, will quickly corrupt into an outdated static snapshot; a seemingly poor document system, with good writing strategies and retrieval tools, will be able to grow.

Two main structural thinking emerges from current practice. One is...Figure structure— Knowledge mapping is naturally good at expressing relationships between entities, and is suitable for storing the highly connected parts of semantic knowledge, supporting multi-jump reasoning and associated discovery. The other one is...Level structure of the file system—Agent is naturally fit to operate a file system, with a hierarchical catalogue and document that can carry multiple memory elements such as the Persona definition, Skills/Rules, an extract of experience and which is fully readable and editable for humans. The two lines of thought are not superior to the other, and selection depends on the characteristics of the memory content: the syntax of the relationship tends to be graphic, operational experience and norms favour documents, while factual records may be more suitable for structured databases. The actual system is often mixed with multiple carriers, and the key is the soundness of the growth and amendment mechanisms on each vector.

Activate and Write: Two Ends of Long Memory

How to activate long-term memory when needed and how to write new content back into long-term memory is two equally important sides of the Memoory system. They should be driven by Agent itself -- not passively responding to queries or receiving writing instructions, but by Agent itself deciding when to activate, what to activate, when to write, what to write. At the same time, both activation and writing effects depend heavily on the design of the storage structure itself: the structure determines what path to recall is feasible and where the information should be placed at the time of writing. And, to put it further, it's not really necessary for acoustics to stop."When to read and when to write"This layer of memory.Organizational structure itselfIt can also be maintained by Agent, rather than organized by a pre-written control stream.

In the early days of the Memoory study, RAG was the core -- memory is about the same as retrieval, with attention being given to finding the relevant paragraphs from the external document. But now RAG is just a component of Memoory. Long-term memory can have different activation paths:

Character Match (Grep/ Glob/BM25): Accurate text pattern matching, fast positioning for known keywords, particularly useful in unstructured data
Semantic Search (RAG): Semantic matching based on vector similarity, specializing in handling vague intent and natural language description, is the core technology for traditional memory retrieval
Structured Message Source (Filesystem/LSP): inherently hierarchical and type-based information search path, Agent can use directory structure for navigation, LSP for code-level sign jumps and references
Knowledge mapping: multi-trip reasoning along entity relationships that is appropriate for answering questions about the link to discovery and causal links

These instruments are not mutually exclusive choices, but correspond to different memory vectors and different types of problems. Text-type memory is suitable for keywords and semantic recall, code and rules are suitable for structured indexing, and highly relevant semantic knowledge is appropriate for mapping. The concern here is not to build them into a set of projects, but to show that long-term memory is not static warehouse: It must be able to be activated when the current mission needs to be and enters Working Memoory with the right particle mass.

And the thing that goes with activation is,WritingThis issue existed in memory studies long before LLM, and was not a retrievable ex post project detail. Writing side is also driven by Agent, and questions to answer include:

What's it say?: extracts memorable points from interactions and actions and filters noise
Where do you want to write?: to judge whether this information should be updated, Persona, settled into experience, Rube, written into the knowledge base as semantics, or compressed into conversational memories
How?: Reweight and align with existing memory, select the appropriate structured form, and avoid redundancy and contradictions
How to Update: new information when it appears, is it merge, subsede, tag conflicts, or multiple versions are kept
When does it expire?: which memory should be downgraded, expired, deleted or stopped by default in case of uncertainty
When?: Trigger writing at the right time, not wait passively for the end of the session

Writing quality directly determines the quality of future activations - if the relationship is not structured, deweighted and formatted when written, and the robust search mechanism can only be filtered in vain from a collection of redundant and contradictory memories. Activate and write together to form the end of a long-term memory, Working Memoory is the operational entry point for this loop into the operation.

Multi-Intelligent Shared Memory

So far, the panorama describes the memory cycle of individual Agent. But when multiple Agents work together, memory and communication become significantly more complex -- They often need to work together to maintain a single identity.Share memoryAnd not just the private memory of each.

The difficulty here is that the monomer problem is magnified by a dimension.Also WriteThe first one is the "Agent" of the blog, which is a blog for the blog, and which is the first of its kind.Program to cross AgentThe credibility of a memory, written by it and based on its observations, and its coverage, is determined.Shared and private bordersNeed to be managed visibly: what memory is in the consensus of the team, what is a local view of a certain Agent, mixed to contaminate the overall state. And...Communication itself is a writing.The news between Agent is not a matter of sharing memories, how to weigh it, how to avoid taking ad hoc consultations as a matter of fact. In the single memory."Literacy"All of the challenges are re-examined in the multi-smart landscape, and additional costs of consistency and coordination are addressed. This paper does not extend the multi-smart body line, but it needs to be treated separately in terms of the evaluation design, and the later version of the assessment blueprint will revert to this point.

Summary

The above is the anatomy structure of Agent Memoory: five types of memory, activation and writing at both ends, Working Memoory as an entry point to run time, and the problem of sharing memory that is amplified in a multi-smart landscape. Agent Memoory has been retrieving library from the outside.Life cycle management of intelligent cognitive status- The difference is not who."There's a memoory."And who knows what to write, when to update, when to delete. So, anatomy comes here, and the next two chapters answer two questions: How this is built up, how it's evaluated.

How did this figure get built?

The panorama on it describes Agent Memoory."The way it looks."- Five memory types, activation and writing at both ends, Working Memoory as the running-time entry. But this picture was not invented in its entirety one day, but was gradually forged in 2023-2026 by several lines of study on continuity of dialogue, learning of experience, context, timing and governance needs. The following is a phased summary of what has been addressed at each stage, which nodes have been advanced on the landscape and what has been the inspiration and gaps left for follow-up.

Phase I: Establishment of basic grammar (2023-2024)

2023-2024 is the window where Agent Memoory moves from piecemeal exploration to consensus. In both years, the cognitive cycle has its first full example, with memory types disaggregating experience and skills from a single dialogue log, and Working Memoory being officially recognized as the time portal for long-term memory, and the evaluation for the first time has helped to stabilize the direction.

The first complete example of a cognitive cycle. If only one task defined the language of the subsequent discussion, it would still be Generative AgentsI'm sorry. It links the memory process of Agent to the whole: to write observations to memory cluster, to retrieve history with relevance, proximity and importance, to form a higher summary through reflection, and to drive the planning with memory and reflection. This is exactly what it looks like."Incoming. Action. Out."The cognitive closed ring: Memoory stream is the prototype of long-term memory, retrieving is the prototype of memory activation, reflection has both the role of writing about extraction and the evolution of Persona, and programming is the reasoning on Working Memoory. It's not just early in its history, but it's providing anchors for all directions. - I'm sorry. The later personalized memory is upgrading the content and retrieval conditions, reflecting on the quality and form of reflection, the image is upgrading the structure of the memory stand, and the long-range Agent is upgrading the planning and skilling use.

Memory type is divided from homogenous observation stread. Generative Ages provides the skeleton, but the memory on the skeleton is homogeneous - all of the entries are written in the same flow format. One of the key advances in 2023-2024 was the gradual emergence of different types of memory from this homomorphous stream, each mapping its different positions in the landscape.

The conversation memories are first to be remembered. MemoryBank Systematization. It places several sub-issues of the long-standing dialogue — the storage and recall of the historical dialogue, the event-level summary, the bodya understanding, the forgotten mechanism — in a unified framework that proves that the memory of the dialogue requires a proactive summary, organization and decay, not just a record of the conversation. Following the quote from the Memoory Bank, the conversation memory actually has an earlier feeder: Beyond Goldfish Memoory (2021) discussed the long-range memory of open-area dialogue earlier, and the long-distance memory of open-area dialogue is a part of the conversation.Long Time No See(2022) Rephrase the persona consistency iconic to long term memory,MemoChat(2023) Support long-range maintenance further with abstract keys. This suggests that the memory of dialogue does not suddenly appear in 2023, but rather that it evolves from one-person coherence to one that is written, indexable and reminiscent.

And the memory of experience is... Reflexion and ExpeL Open up. If it's about memolly Bank, what I told you about, what I learned from my failures is what I'm doing. Reflexion writes that the verbal feedback is a successful re-entry and takes back the next round of missions, and ExpeL sums up the cross-task experience as searchable insight and re-engineers the successful trajectory. This line moves memories from factual archives to behavioural improvement mechanisms - In the panorama, it fills the position directly with experience memory, and the answer is what to do in a similar setting.

Another form of experience is that Voyager and GITM Bring in. The two illustrations are that Agent Memoory is not just used to answer questions, but to build capacity: the Skill Library of Voyager is a typical example of GITM that links knowledge, memory, target decomposition and planning, emphasizing the need for a long-range open world mission. If you focus on the conversation, you underestimate the role of Memoory in code agents, web agents, game agents and embodied delegates. In the panorama, this memory (skill memory) cuts across the boundaries of empirical memory and semantic knowledge — it is both how and what it can do.

The experience memory also has a more focused feeder that deserves to be singled out:Agent Workflow Memory(AWM, 2024) Put"Learning from action"Specificize to yesReusable Workstream (workflow)- The general. It extracts recurrent sub-processes from successful tracks and deliberately does two things:Abstract— Replace the example of a value with a variable name"Get cat food."Abstract as"Search for {product-name} in Amazon") to summarize the sub-processes that are specific and that can be migrated across the mission/website/domain instead of replicating the whole example; andOnline, unsupervised accumulation— to judge the success of the trajectory by using an evaluator, to sum up and write back the memory of the track, and then..."Summary-integration-reuse"The snowball effect can be achieved by just dozens of test samples that significantly exceed the baselines that are not self-adapted (WebArena, Mind2Web, relative success rate up 51.1% and 24.6%, respectively, with fewer steps). In the panorama, AWM narrows the memory of experience to"Reusable program workflow"This particular form, and for the first time, it's a matter of experience.Online accumulation closed loopsReally run up. But it also touches only one side of the story: workflow only increases, is independent of each other and does not deal with conflict abating, failing and governing. This one."Online summary of reusable experiences"And the idea is that the Dynamic Cheatsheet and ACE will be the object of a more generalization.

The entry point for the operation of long-term memory is officially recognized as independent. Generative Acts defaults that all history can be retrieved, and the Memoory Bank defaults limit the content required.MemGPT The first positive challenge is the assumption that working context is not unlimited, that history should not be present at the same time, that memoory needs page breaks, exchanges, wrapping and retrieval -- that Agent needs to manage context and external storage like an operating system. And that contribution to the landscape is fundamental: it turns Working Memoory from an implicit assumption to a visible system node. Before MemGPT, the focus of Memory research was almost exclusively on the side of long-term memory; after MemGPT, how long-term memory entered the current workstation, what particle mass entered it, and when it was returned to external storage became an inescapable border issue.

Memory can be an abstract, as it is constantly rewritten, not just an original log. CLIN It is an exercise that is easily underestimated at this stage. It designs memory into a continuous and renewable source-and-effect abstract text library, emphasizing rapid mission adaptation and cross-environment adaptation. CLIN is inspired by the fact that memory does not have to be just a pile of original conversations or tracks, that it can be abstract knowledge of causality, and that its value lies in continuous rewriting with the experiment, rather than only temporary retrieval in the context of reasoning. This is for later. A-MEM、RMM And when the work of highlighting the evolution of memory is planted -- at the writing nodes of the landscape, Clin for the first time hinted that it was not just an append, it could be rewritten and refined.

The evaluation has made the direction steady. By 2024 LoCoMo And then, for the first time ever, Agent Memoory had matured into the direction of research. It has a very long history of multisession dialogue and has put tasks like QA, event summary, multi-modular dialogue generation into a unified assessment. The historical significance of LoCoMo is not"Another data set"Instead, it establishes the principle that Agent Memoory should not be tested by case-shows alone, but must be systematically evaluated. The L1-L6 evaluation framework, which is presented in the post-assessment chapter (for the full definition of the framework, see the chapter “Assessment”), is largely covered by LoCoMo, which mainly covers callability/activation timeliness (L3) and behaviour consistency (L4), and has not yet touched upon the correctness of writing (L1) and the migration of experience, uncertainty management (L5/L6), but it provides the basis for subsequent upgrades.

What's left of this phase on the landscape? Looking back at 2023-2024, each key task can be mapped to one of the nodes of the panorama: Generative Agens gives the first complete example of the cognitive cycle; Memoory Bank, Reflexion/ExpeL, Voyager/GITM/AWM fills the conversational memory, memory of experience, skills memory and reversible workflow; MemGPT upgrades Working Memory from an implicit assumption to a visible running entry; Clin implies memory can be continuously rewritten on the Agenic writing node; LoComo makes the evaluation framework possible. But this phase also leaves a clear gap: storage structures are dominated by flat-way numerics, time is hardly discussed, writing controls are still inspired, and the assessment is primarily whether they can be answered, not written, and are up to date. These gaps are the direction that the next phase will begin to fill.

Phase 2: Space filled and panorama shaped (2025-2026)

The first phase leaves a clear gap: the boundary between the writing of inspirational, temporal absence, storage of flat-directed vectors, long-term memory and Working Memoory is not systematically addressed, and there is a clear fault between memoory framwork and the real deployable memolly player. The period 2025-2026 has been moving along these gaps, while the memory of parameters has emerged as a new dimension. The table below, which is coordinated by the panorama nodes, summarizes the way forward for each blank space:

Space	Moving forward	Representation	Panorama Node Upgrade
Writing by inspiration	Write modelling as a form / sort / evolve / learnable operation	A-MEM、RMM、MemInsight、Memory-R1	WRITE: + Structure Link Generation + Reflection on Amendment + Auto-Restructuring +Leard CRUD
Time-sensitive absence	User model, → Dynamic image with time line; time becomes first class citizen	Hello Again!、THEANINE、TReMu	Persona: + time line + condition backtrace; search: + time condition
Open the entrance on run time	Level Summary / Gest+recall / Protection-demotion / Application	HiAgent、ReadAgent、Memory OS、Dynamic Cheatsheet、ACE	WM: Compact → adaptive playbook
Storage = flat vector	Time / episodic-semantic Layer / hailarchical schemata / Evolutionary Popping	Zep / Graphiti、AriGraph、HippoRAG、CAM	Carrier"Figure": + time dimension + hierarchy + incremental organization
Infrastructureization/serviceization	Extract - Update pipe / Enduring image / Production access	Mem0	Memoory: Research prototype → deployable memoory player
(New) Parameter Memory	The internal memory of the model becomes the third layer of complementarity.	MEMORYLLM、Titans、MemGen	Parameter Memory: Fake Line →3 Layer View

Key insights by nodes are presented below.

Writing: From application to active crime, then to learning CRUD. The most worthwhile thing on this line is to get it going. A-MEMBecause it turned the line that was buried in the "activated and written" -- the structure of memory itself can be left to Agent -- into a complete system for the first time. A-MEM uses the idea of a card box to write each memory into one.Atomic Note Unit: In addition to the original content and time stamp, LLM field generation of keywords, labels and a context description, with an upward vector embedded, and a set of links pointing to the relevant memory. When writing new notes, the system will first recall the top-k history notes with embedded similarities, then give them to LLM to analyze semantic connections in these candidates.Self-determinationWho should we link to?"Box."(box) and the same memory can be placed in multiple boxes at the same time, more flexible than the traditional card box. The most important step is... memory evolution: New Notes AccessionReverse TriggerRewriting of relevant notes - updating their context descriptions, keywords and labels to allow the entire memory network to continue to refine as they interact. And that just happened to be the seed that CLIN buried -- written not just aspend, but rewritten as an already existing memory. The search side was upgraded accordingly: when a memory was struck, the link in the box was taken out, and the retrieving was accompanied by embedded links and LLM analysis of three clues. Put A-MEM and angstic RAG together, and the upgrade steps are clear: like Self-RAG This job makes agency happenSearchStage - Autonomy to decide when to retrieve, what to retrieve, whether to accept; and A-MEM pushed the agency toMemory structure itself— Organization, links, evolution are driven by LLM, not by preset schema and fixed workflow frames. It removes the rigid control layers of MemGPT / MemooryBank / ReadAgent that require pre-set storage structures, writing times and retrieval paths, and allows memory systems to be developed for the first time."Hand over the package to Agen Management"The prototype. Following the same node:RMM(ACL 2025) Dismantling memory management into two types of reflection;MemInsight(EMNLP 2025) Structuralizes the perspective when it is written: unlike A-MEM, which relies on manually defined notes template, it allows the backbone LLM to be written in a way that is not the same as the A-MEM. AutonomySemantic properties are extracted from historical interactions - both physical dimensions (directors, years, types) and conversation dimensions (user intent, preferences, emotions) and can be marked on both particle scales at the turn or session levels. The properties that are automatically generated in this layer are deposited into a layer of structured annotated discourse, supporting the same memory at the same time."Properties-based structured filtering"and"Semantic Search based on Attribute Embedded", in front of the sidefolding"Activate and Write"The structured SMS and semantic retrieval (RAG) that were discussed separately were harmonized; they were validated in the three categories of dialogue recommendations, question and answer questions, and in the summary of events (LoComo retrieval recall rate was about 34 per cent higher than the RAG/DPR baseline). The blogger adds:Memory-R1 The memory Manager learning to implement ADD / UPDATE / DEPET / NOOP, Answer Aget is learning to model memory management as a learning action space. It is not about the problem being solved, but about the clear route from the prompt rule to the lost policy for the first time.

Persona: From static preference to moving memorry system. The year 2025's personal memory study is a positive challenge."User preferences are stable."This assumption.Hello Again!(NAACL 2025) Proves that individualization is not a simple restatement of history, but contextualized memory development;THEANINE(NAAML 2025) emphasized that old memories did not amount to noise and that it was important to know under what conditions they were still open to interpretation;TReMu(Findings of ACL 2025) is independent of time sequence, duration, updated relationship and validity - none of which can be automatically resolved by ordinary search. The real change in this line is:Starts to see user model as a writingable, compressable, remixable, mapbleI'm sorry. In the panorama, Persona Memory From"Initial setting + evolving preferences"Evolution into"Dynamic images with time lines, traceable trajectories, traceable back to old state under certain conditions"; retrieving side-sync upgrade - time becomes a first-class citizen of the search conditions. To add to this, this line is advancing only by the temporality of the outer memory layer, which is a three-tiered common problem: The whole context layer is the old context is corrupted and context rot, and the external memory layer is advancing, while the parameter memory layer is very difficult to edit -- and the three layers are far from accurate."The memory changes over time."This thing is handled.

Run-time portal: From capacity constraints to online adaptation. MemGPT makes Working Memoory the node of the visible system, but the question of how long-term memories enter the current workstation, how they are compressed, how they are replaced in the advance of the mission remains open.HiAgent(ACL 2025) Level summary and replacement around subgoal;ReadAgent(ICLR 2025) Proof of the current reasoning does not require the full original text, but only the operational gist and precise retrospective capability;Memory OS(EMNLP 2025) Make short/mid/long-term a hierarchy and define four types of operations. The blogger says:Dynamic Cheatsheet The black box LLM can also accumulate external cheatsheet, and then the test can be tested.ACE This idea is systematically translated into a Playbook-style Generator / Reflector / Curator cycle, so that memory is no longer merely cumulative, but can be dynamically updated and forgotten as the system runs, mission feedback and time passes -- If the AWM of the previous phase simply consolidated the success trajectory into a workflow, ACE expanded it to reflectable, addable and modified playbook maintenance loop. The real exposure to these efforts is the boundary between long-term memory and the entry point at the time of operation: only if the memory bank itself is of high quality can it be said that it provides sufficient and inordinate information to Agent.

Storage structure: the chart moves from theoretical options to practical paths. Zep It's a good paper because it's not."Put your memories in the database."So simple. Open source project behind Zep Graphiti The fact that the memory is taken apart into three layers: original news, text or JSON, which is left as an episode, and entities and relationships that are extracted from these episodes, still leads to the recovery of the source. The real point is time. Graphiti writes facts as sides of the duration of the period, distinguishing between the time when the fact occurs and the time of the system ' s ingestion; when new facts enter, it does not simply cover the old facts, but tries to weigh the relevant sides, judge the conflict and mark the overturned edge as invalid. So,"Users like lattes."and"Users have recently started to stop coffee."There is no need to erase each other: the former is still history, and the latter is the present state of recall that is more appropriate. This is also the most interesting place to talk about Zep. It takes memory story from a pile of records that are getting messy, and it moves one step towards a state system that can fix, preserve, and look back in time;AriGraph Combine episodic memory with knowledge-graph world model, making memory both a world relationship expression and an event trajectories record;HippoRAG The project is being organized in a non-eurobilogically inspired manner.CAM Further, map the construction of the construction of the incremental structure of the schemata and retriev the map from the prune-and-grow search"Relationship means carrier"Pushing As"Structural memory patterns for sustainable growth"I'm sorry. In the panorama, long-term memory carriers Medium"Figure"It is not just an option, but a structure with time dimensions, a layer of abstraction, that can be restructured with new experience.

Infrastructureization: from research on prototypes to deployable memory player. Mem0 The point is not to propose another form of memory benchmark trick, but to explicitly package memoory into a separate infrastructure: users need only use simple tools to call without paying attention to the details of the realization of the memory in the back office. Although Mem0 only achieved the simplest form of a Naive RAG, the design of this decomposition allowed for further upgrading to a graph memory on this basis. It has formed a single, heavy, updated and re-organized four actions into a memoory player accessible to developers in many studies, representing the critical step of Agent Memoory from an experimental structure to a production-level service.

Parameter memory: Complementarity third layer. There is also a parallel spectrum in 2024-2025: back to the model. The cost-balance of this line and the course of training (why not to go straight, how to synthesize data and then train) have been carried out in the initial three-tiered framework, where it is marked only in the historical context.MEMORYLLM (a) Continue writing in the reasoning period with a player-wise memorypol, proving that memory can be updated late substrate;Titans Rephrase the question to"How to learn to remember and forget in the best time"The core is no longer a search, but a search;MemGen Further, re-formulate the current need for meta-native space memoory in the reasoning by trigger + weaver. Together, they fill out the three-layered picture: external management memory, working memory, current workstation, model-international ability memory, and settling the repeat experience into lower-cost sustainable adaptive capacity -- – Parameter memory is always complementary rather than a substitute, and the main structure of the panorama (non-parameter memory is the main line) remains valid.

This phase leaves new gaps. The initial Learned Policy (e.g., Memoory-R1) has emerged from writing controls, but is far from universal: the motion space is still limited, the training signal and scene coverage is still thin, and the multiple types of memory and governance needs are not covered. Conflict mitigation and selective forget-outs still lack mature programmes, and the auditability and negability of internal memory of models are far from being resolved; while Dynamic Cheatsheet/ACE has given a clear line to the evolution of the online context, the communication control, reliable reflection and long-term assessment coverage remain inadequate. The whole picture is in place with bones and flesh, and the next challenge is to keep the system operating in a sustainable, credible and manageable environment.

Evaluation: Why can't we keep up with the Memoory evolution?

The skeletons and flesh are in place, but one difference is becoming more apparent:The methodology has been developed from"How?"Push to"How? How? How?", and the subject of the evaluation still stays"Can you call it back?"。 If the assessments fail to keep pace with the evolution of the methods, it is difficult to know whether the systems are progressing or repeating the same things in a more timid way. The chapter brings together the full-text evaluation discussion in one place: first, to establish an assessment framework, then to see where the available benchmarks are measured, where they are missing, and finally to try to design a truly systematic memory assessment.

Evaluation framework: L1-L6

To judge whether a memoory system is okay, a set of dimensions must be in place. Put your memories back."Form-activate-Action-Correct"In the life cycle, the assessment must cover at least six levels:

Level	Assessment dimensions	Core issues	Typical failure scene
L1	Writing Correctness	It's written in the right place, right?	Use user jokes as a preference for Persona.
L2	Update correctness	I've got a memory correction, right?	User changed address, old address not covered
L3	Call / Activate Timeliness	Did you use it?	You've stored user allergies, but you're not calling them. Back
L4	Consistency of behaviour	Does the behaviour match the current memory?	Remember the user preferences and the conflicting responses.
L5	Experience migration control	Is long experience leading to manageable generalization?	A Project learning normative errors applied to B Item
L6	Uncertainty treatment	Does it leave abstain and marks when it's not sure?	Forced fabrication rather than marking uncertainty when memory is blurred

A simple alternative.: If you don't want to break it down, you can throw it to a fixed, complex environment with end-to-end tests -- can you remember and do something about it? But the end-to-end test only tells you that."Okay?"I'll tell you when I'm done with L1-L6."Where's the wrong part?"。

Core assessment and what they're measuring.

With L1-L6, this ruler can see where the benchmarks are now moving:

Benchmark	What are you talking about?	Corresponding assessment level
LoCoMo（ACL 2024）	Overlong session recall of conversation, summary of events, time/cause understanding	L3 Call / Activate Timeliness, L4 Behavior Consistency
LongMemEval（ICLR 2025）	Multiple session reasoning, time reasoning, updating of knowledge,	L2–L4, part L6
HaluMem（2025）	Writing/update/calling three-stage memolly hallucination operation level diagnosis	L1 Write correctness, L2 Update correctness
LifelongAgentBench（2025）	Experience migration, negative migration, skill return	L5 Experience migration control

The evaluation targets have been significantly upgraded in the past year: From recall(can you answer) Go trustworthy write / update / use"Way to go, right to go, right to go." stream-level adaptation(Is there a continuing positive migration in the long-term task stream? The system doesn't just need"Remember."And there is a need to write correctly, to refuse more steadily and if necessary, and to continue to generate positive migration in the long-term task stream.

L1-L6 Coverage and Gap

Level of assessment	Overwrite	Still missing
L1 Write Correctness	HaluMem (first time locate memory hallucination to the operating layer)	This is the only one that lacks a variety of scenes.
L2 Update Correctness	HaluMem、LongMemEval	System-level diagnosis of conflict resolution remains weak
L3 Call / Activate Timeliness	LoCoMo、LongMemEval、MemBench	Cover the fullest level
L4 Consistency of behaviour	LoCoMo、LongMemEval、TeaFarm	Better cover.
L5 Empirical migration	LifelongAgentBench	Simulate task stream only, without real long-term assessments
L6 Governance capacity	LongMemEval（abstention）	Remove / rollback / privacy / basic blank

The current assessment is dense on L3-L4, just starting on L1-L2, seriously inadequate on L5-L6-- – The forward direction of methodological research (in terms of control, governance) is where the most inadequate coverage is assessed.The most important concern of the front-line research is precisely the assessment of the most deficient.

A systematic set of non-IID memory assessments.

L1-L6 tells us what dimensions to measure, and what Benchmark is telling us is what we need. But there is a more fundamental methodological premise that is often ignored:Memory assessment must be non-IID - it takes into account the functioning of the entire memory system in multiple rounds of interaction, not a set of separate QA questions. Traditional assessments treat each issue as a stand-alone and distributed sample, giving a fixed text, a one-off search; but the essence of the memory system is a state: it determines what to write, what to sink, what to update, and what to answer after these decisions shape it. It is therefore right to allow memory systems to work on their own in a continuous, interdependent, interactive trajectory, with the assessment point based on the state of memory that they had accumulated — the previous round being misrecorded and the latter being affected. That's how it's gonna work."System-level memory capability", not"I'm not sure if I can get a good search on the regular text."I'm sorry. In this context, a comprehensive set of memory assessments should cover the following seven directions:

Correct answer to memory irrelevant (context rot baseline): After injecting a large amount of memory or a very long context, a number of questions are first measured that are completely unrelated to memory. If these questions that should have been answered steadily begin to deteriorate, then the context is already"Corruption"— It is the foundation of all capabilities, the foundation of all capabilities, which has collapsed, and the recall of the latter is meaningless.
The correct recall of reference memories: Whether the content of the external database or previous dialogues can be retrieved accurately if needed. This includes both classics of need-in-haystack.Learn from experience during testingCapability - Is the system capable of immediately settling the feedback from the current interaction and reusing it later (Dynamic Cheatsheet/ACE, as previously mentioned)?
Rect MultisourceQuestion Answer: To connect the memory system to a question and answer process that requires multi-step reasoning and multi-source integration, to see if it can integrate information from different sources,Especially when multiple sources are in conflict with each other.— Is conflict identified, time/credibility compromised or is it indiscriminately confused. This is a direct re-enactment of L2 and a litmus test of conflict mitigation.
Lifelong learning and memory compression: Accumulation of memory on a long stream while examining whether compressed memory is still retrievable. The hard part is to design a measure at the same time."Continuous absorption"and"Compression is not defunct"Mission - Memory needs to grow and not lose critical retrievable information after compression (responding to the post-ReadAgent gist memoory, the Memoory OS layers, and L5).
The memory of the cross-scenes fades.: Compare all Agents with memory modules with memoory-on/ memoory-off values to quantify the exact value of the memory module using the difference of the final effect. Broad coverage: role-playing, chat robots, personal assistants, game-type Agent (e.g. Minecraft, echoing Voyager / GITM), code generation (e.g. SWE-bench). Absorption is a test of memory."It's not working."The most immediate means - if the memory of the switch has no effect on the outcome, the memory system is set.
Intelake and cleanup, respectivelyThis is the most critical and neglected. Low quality memory for the final system isCatastrophe.And the wrong memory is recalled over and over again, worse than no memory. Thus, inclusion and clearance must be assessed separately as two separate indicators, rather than mixed in the recall rate. The cleanup can be done fromEnd-of-pipe utilityIt is established that if a memory is actually used but then used, the result is worse, it is proof that it should not be left. This jurisprudence applies equally to empirical memory (operational experience, rule) and knowledge of language/semimanual knowledge (facts, relationships).
Multi-Intelligence Symphony: The above-mentioned reference to shared memory of multiple intelligence bodies is more complex than a single body, and the evaluation is also designed specifically to measure consistency in the sharing of memory under simultaneous writing, the ability to pursue cross-Agent policy, the reasonableness of conflict decisions, and the trade-off between communication costs and the quality of memory.

In order to bind these seven, a set of qualified memory assessments should uphold three design principles:Hold on, no-IID(As a state system, allowing memory to sink itself in multiple rounds of interaction and to continue to explore the tracks of the previous state, instead of using a set of separate QA questions)Separate intake with cleanup"It's different to include it, right?Cross-scenes to make a difference.(Perform the true value of memory by using the utility margin). These three are just the right ones for the first one of this chapter -- the way it's written, the way it's changed, the way it's governance, the way it's measured, the way it's measured, not the way it stopped."Does it look like it?"。

Next: the grid still empty on the landscape

The gap in assessment points to gaps that are not filled by the methodology itself. The skeletons and flesh of the map are in place, but the challenge of determining long-term availability is still far from being resolved. If one by one is to follow the nodes of the panorama, the most worthwhile open-ended issues can be grouped into five categories:

Write policy is to move from inspiration to more robust and transferable learning strategies.Memory-R1 It has been proven possible to learn ADD / UPDATE / DELETE / NOOP, but today ' s Learned policy is still in its early stages: limited mobility, scarce training data, cross-mission extension and governance constraints are not yet well modelled. The difficulty is not to prove that it is “learningable”, but to keep it stable and manageable in multiple types of memory, multi-source conflicts and real long-term interactions.

Conflict absconding and selective oblivion have not been fully defined. Old and new preferences, time-state changes, multisource memory consolidation all point to the same issue: memory needs to be “modified” and “deleted” not only for “addition”. A more mature definition of the question is needed as to why, how to validate the deletion, whether to allow rollback and how to meet privacy and compliance requirements.

Time validity needs to be the default field. High-quality memories store is likely to support timestamp, Validity interval, supported-by/contradicted-by, time-aware retrieval and stale memory. The static preference of “like latte” and “recently beginning to stop coffee” cannot be considered as equal rights, and memories must take time, version and validity. As has been repeatedly emphasized, time is a common dilemma that cuts across the whole context, external memory, parameter memory — it has not been really resolved at any level.

Credible governance is the furthest from production. An deployable person needs at least the capability of provenance (source tracking), confidence/uncertancy label, stale/confect action, operation-level write/date audit, poisoning corporate, and rad overwrite/ unlearning.

Long-term assessments of the real world are still scarce. Most of the current benchmark is still simulated, time-consuming compression. The missing items are real interactive assessments from the months to the weeks, more natural multi-application switching to correct errors with humans, operational logs and audits of write / update / delete, and STread-level indicators for revision / forward transfer / backward inversion. This is the direction that the blueprints for the previous chapter were intended to add - adherence to non-IID, integration and clean-up of separation, integration of trans-situation and consideration of multi-smart synergetic memory.

These five types of questions can be judged in one way: if you see angent as a system that requires long-term online work, memoory is not an option, but a prerequisite for it to become a long-term intelligence. A stronger anent is not just a longer window. It would explain what should be written, what should be activated, what should be updated or deleted, what should be preserved, what should be internalized into adaptive capacity, and why a memory is credible and when it is not working.

And putting these together can also show a direction: instead of using a pre-written control stream.External organizationMemory - MemGPT-style break-up, MemooryBank-style fixed frame - rather than moving on along the lines of A-MEM, so that the reading and writing of memory, organization, links, evolution and governance are all taken together Agent SelfTo take responsibility: to regulate these acts with hints and Skill, radicalize even a memory, Agent, free the entire memory system from rigid control structures. It's a design orientation that needs to be added: instead of replicating the memory compartment of the human brain, the structure should be based on the following:Organization of the document systemMoving forward - Agent is naturally well-equipped to operate files, directories and links, readable, editable and directly indexable by tools, which is more smooth than pushing memory into a humanized stratification model, and is closer to the engineering of the matter of "making memory systems themselves into intelligent bodies".

Concluding remarks

Looking back at this whole connection, Age Memoory is not just a "back-to-back search library." It is moving from a search question to a life cycle management issue for long-term cognitive states: how memory is formed, when activated, how action is refined, how old states are corrected, how conflicts are managed and how failed content is removed. This paper starts with three layers of context, external memory, parameter memory."Where do you remember? What do you mean? Fine."Clear — it is essentially a trade-off between the costs of reasoning and training, and external memory is the most engineering option for the moment; it goes along the path of anatomy, history, evaluation, and ends up in a blueprint of assessments that adhere to non-IID, separation, integration and clean-up, and cross-scenes.

So, the point of Memoory’s competition is not to have a memory library, but to “write right, change right, be activating when it is time, be governed when it is time to fail” and whether we can design a real assessment of these differences. If this judgement is established, the next phase of Agent Memoory is not simply a search project or a mere context management, but rather a long-term memory that is properly formed, continually modified, activated and credible.