Agent Memory and Runtime Atlas: External Memory, Runtime Systems, and Framework Memory

Writing status: This is the place-in draft. Only the article location and title pits are kept, the product index is not expanded and the long content count is not made.

The questions in this article can also be addressedFrom memory formation to memory governance: the panorama of Agent Memoory、Context is All You Need: Context Project for Smart BodiesHow the concept of a relatively close read together is developed in different contexts.

Agent Memoory and Runte technology inventory: hang memory, running time research and capacity-building in a framework

The article no longer follows the L1/L2/L3 mapping method and does not rank as a product. The real answer to follow-up is how a system is remembered, how it is taken, how it is modified and how it is forgotten. Here we use the example of a real-run system to explain how a complete Agent Runtme, a Memoory Framework and Agent's development tools help users manage context and long-term memory. This will be a case study of the previous discussions on Research and Context Engineering.

Emory Products and Research (in position)

Mem0 / OpenMemory

Zep / Graphiti

Letta / MemGPT

TencentDB Agent Memory

Cognee

Supermemory

MemOS

Generative Agents

Reflexion

MemoryBank

A-MEM

LongMemEval

HaluMem

MemoryAgentBench

MemConflict

LongMemEval-V2

Agent Runtime Research and Products (in position)

AIOS

SWE-agent

OpenHands

Agent S

OSWorld / BrowserGym

Magentic-One / Magentic-UI

Claude Code

Codex CLI

OpenClaw

Hermes Agent

Pi.dev

Agent Framework provides memory capability (in place)

This section discusses the memory capabilities that Agent Framewerk provides, not the specialized Memoory Layer, nor the personal memory that model manufacturers make in their products. In other words, what concerns us here is how much memory it has addressed for you, and how much complexity it has left to developers, when the developers have chosen a framework.

I'll break it down five dimensions.Short-term and CropSee how the session history enters the context window and whether the framework provides cut-off, summary, command or header.Long-term memorySee if history information can be used to repeat itself across sessions, or only provide an interface to the extremity provider.User MemorySee if there are any user, session, app, agent, that separates the fields from the memory of different people, different missions, different product entry.Work memorySee if the framework supports taskboard whiteboards, TODO, intermediate decision-making, structured operational status, rather than simply stuffing everything into chat records.RAGIt's about how it's going to retrace external knowledge, and whether it's using RAG as a knowledge base, using Memoory as user experience, or simply mixing them into the same subject injection.

These five dimensions are not meant to score each frame, but to expose their memory philosophy. They can all be called memory, but the complexity of the work that they are undertaking is completely different. Again, this interpretation will facilitate your choice of the right Framework, based on your needs and the degree of alignment of the framework.

LangGraph / LangMem

Short-term and CropLangGraph's short-term memory is not just a fat buffer, but a thread-level graph state.checkpointer Saves the same messageages, tool result, intermediate state, human-in-the-load stop point in the thread to allow graph to recover, try again,time level. Cutting is not just about deleting old messages, but deciding what messages are left, what messages are removed, what history is summarized, which nodes remain engaged in reasoning before status goes into the model. State is a status dictionary for the status machine that can store short-term conversation memories with messageages, but the specific strategy is decided by the developer. LangMem has completed the upper layer of this type of containment, reducing the development of direct handwritten message-writing logic.

Long-term memory: Long-term memory storeI'm sorry. It crosses theread exists, saves JSON documents, each memory has namespace and key. You can save user facts, preferences, mission experience or application-level knowledge by namespace, but namespace, schema, privileges, forget and merge rules are all designed. LangMem adds precisely “How to write memory” and “How to recall memory”: angent can actively preserve and search memory in hot path with memoory tools, or regrace memory manager in the backstage, remove weight, merge, update and maintain memory consistency. This layer is very close to the real memolly player that the Agent product needs.

User Memory: LangGraph does not define you UserMemory Schema. User memory is usually long-term. store A namespace design, for example, by using user id, app id, tenant id, project id combination, and using key to distinguish between programme, provision, episoc memory or mission experience. The frame gives you a thread storage and recall, without deciding what the user image should look like.

Work memoryThis is the strongest piece of LangGraph, and it's what we've already mentioned in short and short cut. grap state is natural for saving mission plans, intermediate decision-making, approval status, nodal output, tool results and failure recovery information. It sees the work memory as a sustainable memory performance rather than a temporary text in the context of the model. This is critical for long missions, multi-step processes, human clearance and operational recovery.

RAG / Store: LangGraph store Long-term memory can be used, as well as vector-retrieving and external knowledge. RAG is not an isolated plugin here, but a call for nodes, states and tools in the graph: the search results can go into state and the subsequent nodes continue to be processed. LangGraph/LangMem is the most similar group of Memoory infrastructure in the group, which is organized so that it is difficult to design the system by the developers themselves.

Mastra / Agno / AgentScope / CrewAI

These four frameworks can be seen together, not because they are not important, but because they and LangGraph deal with the same set of issues, but with an abstract hierarchy. LangGraph, put state、checkpoint、store Exposure to bottom infrastructure allows developers to define their own systems for memoory runtime; Mastra, Agno, AgentScope, CrewAI is more like putting some systems directly into a frame, allowing developers to get them to the available form more quickly. The difference is not whether there is Memoory, but whether it is in the framework or in the application developers.

Short-term and Crop: Unlike the gramp state of LangGraph, these four are more like a frame of capacity with a short-term context. Mastra follows the TypeScript adjunct route, message history binding long / source, memory processes filtering, cropping, re-ordering or summarizing the context before the model is called, avoiding developers writing a bunch of messages themselves to clean up logic. Agno also splits the session history, which is responsible for “what was just talked about”, and the session sessions, which is more appropriate for continuous dialogue in products. Memory module saving for AgentScope Msg, and tag messages with mark, RectAgent also provides memory control, and addresses token limit. The short-term context of CrewAI is embedded in the execution of Agent, Task, Crew, Flow, and the developers are more involved in configuration of collaborative structures than in the maintenance of a state condition.

Long-term memory: Long-term memory of LangGraph is universal storeSo it's like a database; these four are more like a product-based memolly player. The most obvious feature of Mastra is:semantic recall The blogger says that the government is not a party to the law.observational memory The main goal is to use the backstage, Observer / Reflector, to organize long conversations into a more detailed observation log, targeting the small-sized Raw history, while retaining long-term experience. Agno's. Memory / MemoryManager More directly, create, retrieve, update, delete long-term user memories and allow ant to manage memory by themselves through memory tools. AgentScope has a long-term memory interface, which is also connected to Mem0LongTermMemory, and ReMe, a more research-oriented and ecological memory project. Crewi, the most recent direction is unity. Memory Class, using LLM to infer coope, category, importance, synonym similarity, recency and importance at recall, and instead try to integrate them into a unified API.

User Memory: The most interesting contrast here is Agno and Mastra. Agno clearly distinguishes between user memoory as a user fact, preference and context learned from the inter-conference, and the boundary is very productical. Mastra saves multiple multiple hits using a sustainable Markdown status block, using source scope, or limiting them to single jobs, suitable for writing user images, long-term objectives, preferences and current job summaries. CrewAI's unfied memory can rely on scope tree and LLM to analyse the user or project memory, but this automatic organization is convenient and means that you have to check carefully whether its boundaries meet product rights requirements. AgentScope also saves preferences and facts, but it is more generic: message mark, session state, long-term memory, and user image schema still needs to be defined.

Work memoryThe work memory of LangGraph is the graph state, so it is well suited to consider the plan, approval, nodal output, failure recovery as a sustainable state of implementation. Mastra and Agno workspaces for more finished products": Mastra working memory can act as an active scratchpad, workflow state that saves process steps and intermediate products; Agno uses session state, Team, Workflow, AgentOS runtime to carry tasks and multiple agent collaboration. The working memory of AgentScop is more like running-time state management, and the system program, memoory, constext, tool, session is integrated into state / session management, suitable for experimenting with multiple angents, tool calls and re-runable. CrewAI is the most collaborative model: the work memory is hidden between Agent / Task / Crew / Flow, like the status of the project division of labour and process advancement, rather than a whiteboard.

RAG / Store: LangGraph sees RAG as node, state, and tool call in the graph, so the search links can be organized in full. The RAG colour of Mastra is mainly supported by the vctor store for security recall: history is embedded, retrieved, and recalled with the following window. Agno separates Knowledge, and it is clear: Memoory handles user and session experience, and Knowledge handles external knowledge base. AgentScope has a combination of RAG, toolkit and long-term memory, using a toolbox similar to Langgraph. CrewAI also has knowledge, history and external tool connections, but its features are uniform. Memoory can be used alone or can be hung on Agent, Crew, Flow. The overall trend is consistent, but access paths are influenced by the backbone design.

So these four can't simply say that they are weaker than LangGraph. The value of Mastra is that the top of the top of the top five is that the top five of the top four, which is the most important part of the world, is that Mapescript’s ecological system rarely links the message history, security recalculation, working memory, observational memory, memoory procrastors into a relatively complete set of relatively complete stories. Agno's value lies in the clarity of the boundary, the integrity of the product, the quality of the product, the quality of the product. The value of AgentScope lies in engineering and research ecology, which is suitable for multi-agent, state management, RAG, long memory combinations and direct service delivery. CrewAI is valuable in the collaborative paradigm and harmonization of the Memoory API, which is suitable for fast organization around roles, tasks, teams and processes. LangGraph / LangMem is more like a foundation for memoory runtime, stronger than a programmable, fungible, manageable; these four are more like having some of your pipe shape, faster, more product-oriented, more appropriate for an online service.

AutoGen/Microsoft Agent Framework/Semantic Kernel

Microsoft Agen's family route is a bit messy: AutoGen, Semantic Kernel, Microsoft Agent Framework can speak Agent, but they represent three generations of different abstractions. AutoGen is more like a research-type multi-intelligence framework, Semantic Kernel is more like a business integration and RAG framework, and MF is when Microsoft tries to bind and run the unified operation. Microsoft is indeed a non-savory company that likes to compete with itself.

AutoGen

Short-term and Crop: short-term memory model_context。BufferedChatCompletionContext Keeps the latest N message;TokenLimitedChatCompletionContext (a) The information is removed from the middle when the budget is exceeded, as the context of the budget control of the token is used;HeadAndTailChatCompletionContext Keeps the earliest head and the nearest tail, and uses the sskiped message to mark the history that has been skipped in the middle. These strategies address the context window budget, and there is little compression or cut-off.

Long-term memory: Long-term memory passes Memory Protocol hanger. The basic link is: the packaging of facts, texts or other contents MemoryContent Write memoory story;does it before deduction by memoory based on current context or query query;readoption update_context Write back the recall results to the context of the model. The need for more long-term memory is the use of addI'm sorry. AutoGen only sets this set of interfaces and does not specify specific workflows at the back end.

User Memory: AutoGen no UserMemory Field. User preferences, identity facts, cross-session information available ListMemory Saves, and can also be placed in Mem0, Redis, Croma or custom backends;ListMemory More like an example level of achievement: saving content in chronological order,update_context , and then add the memory to the context. When we're actually developing, we need to use a long-term memory-like approach. UserMemory Problem.

Work memory：TextCanvasMemory It's more typical of work memory. It saves a constantly changing document or canvas, similar to todo.md, lets angent edit long text, draft or schedule with the middle state external, not automatically extracting user facts. His way of work and his classics. todo.md Basically the same.

RAG: In summary, Autogen does not provide its own RAG design, but just another set of interfaces, where you can freely choose and organize your own RAG system. The blog is a different story, and the blog is a different one. They can be recalled with embedding, keywords, metadata or external services, but in AutoGen it seems to be all in the same mode: memoory decides how to search, sort, filter, then pass. update_context The results are injected into the context.

This is Autogen, which is now completely out of date and not worth using, and we can see that he's basically the same thing that was done in the Memoory.

Semantic Kernel

Short-term and Crop: short-term memory of Semantic Kernel ChatHistory and AgentThread。ChatHistory The first is the flow of information actually seen by the model.AgentThread It is abstract (service-side thread id). SK has ChatHistoryReducer This line allows for the breaking of messages and the summary of old messages. This capability works primarily at the Chat history level, and the developers are required to decide when reduce, which system/tool/funct-call structures are retained, and only the appropriate tool envelope is provided.

Long-term memory: Long-term memory is mainly passed Mem0Provider And we'll talk about it at the RAG. It extracts the memories from the thread, followed by recall before the current request. Long-term memory here is not a raw chat archive, but a reusable fact after it has been extracted. SK hangs external memory capabilities on ant-call chain through provider. But at least it does not provide only abstraction and solutions.

User Memory: Semantic Kernel still uses similar Mem0Provider The programme, however, has been categorised.UserId The first of these is the fact that the Internet is a tool for the development of the Internet, which is a tool for the development of the Internet.ThreadId It can be confined to individual tasks. It separates “the person is so long” from “the task is so temporarily”. It's clearer than AutoGen that the SK at least exposed both fields to user/thread, and the memory system itself remains dependent on external Provider.

Work memory：WhiteboardProvider It's similar. todo.md The product provides a fairly good cover and open-box experience. It draws from the dialogue the requirements, proposals, decisions, actions that truly stabilize the structural state of reservations during the process of the preservation of the mission. Even if the Chat history is cut off, angent knows what has been confirmed and what to do next.

RAGRAG is the place where the SK matures. Victor Store is an external knowledge base.TextSearchProvider Retrieve and inject the context of the agent. It's not just a memory device like AutoGen, but rather a more complete connector design around the enterprise knowledge base, vector storage and search program. And this is also a few complete systematic designs, I think.

This is Semantic Kernel: It's not a unified Memoory Runtime, but a group of providers that combines the capabilities of the thread, the history, Mem0, the Whiteboard, and the Victor Store. It is closer to the available enterprise RAG and taskboard than AutoGen, but it is still a combination of capabilities, not a unified memory system, as a whole.

Microsoft Agent Framework

Short-term and Crop: Microsoft Agent Framework short-term memory AgentSession and ChatHistoryProvider。AgentSession Saves the status of the session for dialogue reuse and recovery.ChatHistoryProvider Managing historical sustainability and context management: When responsible for the configuration of the context and the context, reducer and control are compressed into a context budget that can be continued in deduction. Developer needs only a simple configuration.

Long-term memory: Long-term memory is unified AIContextProvider Back. Provider can search for context before calling, or extract memories from new messages after calling. The mechanism here is exactly like the SK, and we can make a proper scope, and then we can make it... AIContextProvider According to the corresponding scope recall, this is the importance of metadata that we have mentioned in other blog.AIContextProvider Nor is it bound to the fixed back end.

User Memory: User memory splits by scope AIContextProvider Achieved. App, angent, user, session can all be borders, avoiding different user and task strings. Although user memory systems are not listed separately, MF has largely found a sophisticated abstract that provides a relatively stable memory layer, but still requires a lot of developer work, and backstage provider needs to achieve complex memory storage and recall logic for front-end simplicity.

Work memory: MAF does not open the box as SK WhiteboardProvider does, using structured working memory. It provides the floor blocks of the session state, context program, workflow checkpoint. You can use them to achieve the whiteboards, mission plans, intermediate decisions and TODOs, but the developers must design them themselves, drawing on the schema, updating strategies, conflict management, and when they are injected into the context. It is rather unfriendly, and work memory is too common and duplication takes time.

RAG：TextSearchProviderResponsible for external knowledge recall. They can be automatically injected into the context or used as a tool by angent on demand. The full inheritance of the mature outer knowledge base programme offered by the SK.

This is Microsoft Agent Framework: It tries to bind AutoGen's multi-smart body heritage and the SG's enterprise RAG heritage to a set of sessions, headory, context programr, RAG, company. It is true that the whole is more mature.

Dify / Coze / n8n

Dify

Short-term and Crop: Diffy's short-term memory is mainly in the context of the session at the Agent and LLM nodes. The Memoory parameter for the Agent node is essentially TokenBufferMemory, the developers can control how much historical information is kept, and the larger the window, the more complete the context the model sees, the higher the token cost. LLM node can also enable memory, and pre-user-assist messages can be spelled back as context. This is about multi-wheel continuity in low-code applications, not a complete Agent Runtime memory system. Diffy will automatically process the system to avoid the context explosion, and of course the classic cropping + compression idea.

Long-term memory: Diffy defaults no real inter-sessional long-term memory. The LLM node is node-specific and does not last between different types of conversion. To keep user facts, historical preferences or mission experience for a long period of time, developers need to update their own logic with external databases, knowledge banks, plugins, API nodes or variables. Diffy provides an application programming platform, which is not a framework for automatic extraction, consolidation, oblivion and governance of long-term memory.

User Memory: The closer to user memory in Dify is the Conversion Variables. It can keep a state in the same chat session multi-cycle dialogue, such as user selection, table fields, task phase, temporary preferences. But this field is still a session level, not a natural one. UserMemoryI'm sorry. If you want to do "what this user has long liked," "what this customer has bought before," you need to get the variables and external storage through, and you need to do it with the memory of the long term.

Work memory: Diffy's working memory is more like the workflow variable and node state. Variables give value nodes, subdivisions, knowledge search results, and tool returns can together form the middle of a task, but it is not a structured whiteboard like SK Whiteboard, nor a user chart like Mastra working memory. This is the program memory of the Workflow, where the schema, the updated strategy and conflict management are scattered in the workflow design.

RAG: RAG is one of Diffy's most mature directions. It places the knowledge base, retrieval, re-ordering, LLM nodes and applications on the same platform, which is very friendly for business questions and answers, for helpers, for forms and for content generation. In Dify, RAG is more external knowledge entering context, and Memoory is more session continuity and variable status, which can be combined in the workflow without a unified long-term memory model.

That's Diffy: it's good for delivery, not for long-term memory mechanisms. It has finished the most common knowledge base, tools, variables, distribution and operation backstages in Agent applications, but the core issues of memory, particularly cross-conference user fact-taking, conflict consolidation, obsolescence, deletion and governance, are still to be filled by developers themselves.

Coze

Short-term and Crop: Coze 's short-term memory is mainly in the context of Bot dialogue, variables, workflow nodes and platform hosting. As a product-based Agent plant, it is more concerned with whether users interact continuously in chat portals than with how to get every message into the context of the model directly under the developer. Context tailoring, historical fusion and nodal state are hidden from the platform, focusing on product design without concern for those details. Of course, the backstage is the one that cuts + compresses, preserving the memory of a single session, which is the most cost-effective option.

Long-term memory: Coze's long-term memory is not pure black boxes, but it exposes product configurations rather than complete memory runtime. Developers can open long-term memory on smart body layout pages; when it is opened, the system automatically extracts user images, user memory points and user manually edited information from the dialogue and writes into the Platform system database. Developer can also configure whether to "support the call of long-term memory in Prompt": when it is open, users can recall memory by asking questions in regular conversations; when it is closed, long-term memory can only be recalled through the long-term memory nodes in the workflow. The new version of the Memory Library also supports the binding of several low-code intelligence and workflows, and uses user UID and channel ID combinations as a separation mark. This is his exposure to long-term memory, but the exact work of the backstage is unknown.

User Memory: Coze user memory boundaries are clearer than Diffy. The new version of the long-term memory is separated by UID and channel ID, and data from different users and channels in the same memory library will not be mixed. Memory content is also classified internally: user portraits, user memory points, user-generated information; in which user-generated information is given higher priority and is given priority when content is automatically extracted in conflict. Developer and debugger can be in memoory > Long-term memory Page views, edits, deletes or emptys your long-term memory, which users can also modify or record through the natural language system. The nodes of long-term memory can be used to recall long-term memory. Coze has a similar ability, considering that we put databases and variables in the user memory category in Diffy.

Work memoryCoze's work memory is mainly in the workstream, variables, databases, knowledge base and plugin organization. It can carry task status, form information, intermediate results and process node output. This is how the workflow is organized, and there is nothing to add.

RAGCoze also has a knowledge base and plugin system, and RAG is here to help Bot to answer business knowledge, call external capacity, and perform product entry tasks. Its relationship to long-term memory is more product-oriented: the knowledge base preserves public knowledge and long-term memory preserves user-specific information. The product is clearly divided, but it does not integrate knowledge base, long-term memory, variables, databases into a programmable memolly runtime. You can only use them through visual nodes.

This is Coze: it's not a code-first factory, but a product factory. Its long-term memory is more like a true user-oriented, personal memory than Diffy, but it is also more transparent in engineering and platform-based. You can quickly make a user-remembered Bot, but not necessarily carefully control how the memory system is written, consolidated, audited and migrated.

n8n

Short-term and Crop:n8n short-term memory is mainly achieved through the memoory subnodes in AI cluster, typically Simple Memoory. It saves the chat history of the current session that can be configured to length, so that AI Agent or chain can know what the user said in previous rounds. This ability solves the continuum of dialogue, a visualized chart history buffer. Still cut and compressed, as in all three workflow platforms.

Long-term memory:n8n does not have a complete long-term memory governance system built in, but it is very good at connecting external memory services to processes. In addition to Simple Memory, n8n can also use memory services such as Redis Chat Memory, Postgres Chat Memory, Zep; and Chat Memory Manager if more complex messages are needed to load, insert, delete and compress. n8n considers memory as an external node that can access workflow.”

User Memory:n8n user memory boundary depends on session key, external storage and workflow design. You can use different sessions id to distinguish between different users, or you can write user information into databases, CRMs, tables or vector banks, and then take it back in the follow-up process. But this is not Diffy/Coze's user image memory for the end product, and it's not a frame level. UserMemory Field. User isolation, privileges, life cycle and removal strategies are all the responsibility of the workflow designer.

Work memory:n8n is really strong at work stream execution. The trigger, nodal input output, condition subdivision, loop, error processing, external SaaS connectors together form its working memory. It can naturally preserve “the step to be taken in this automated process, the system to be returned and the service to be called next”. This is how the workflow is organized, and there is nothing to add.

RAG:n8n can be used to spell RAG processes through vector bank nodes, search nodes, tool nodes and AI Agent nodes. The RAG here is still a workflow assembly: you connect data into vector banks, retrieve, reorder, generate these steps, and it can run. It is not a unified memory runtime, nor will it determine the priority between external knowledge, chat history and user memory.

This is n8n: It's good at embedding AI into automated processes, not evolution into complete Agent runtime. Its memoory ability is practical, but essentially nodal session memory and external storage connection. It works very well for the process automation + a little smart; it still depends on developers for complex long-term memory, user image governance and multiple Agent work memory isolation.

LlamaIndex

LlamaIndex is also relatively special. It is not a typical rag dev trade, but rather a RAG / Context Engineering framework: to organize these external data into the context that can be used in the model. It's got anent, workworkwork, tool calling, but most of these capabilities are about how to bring the right knowledge into the model, and if we don't need Agentic RAG, then LlamaIndex doesn't even have to develop these Agent components.

So LlamaIndex's Memoory is more like an extension of RAG's ability. It saves recent conversations in short queues, and after exceeding the token budget, it can give old news flash to the MemoryBlock；StaticMemoryBlock Putting the background on the table.FactExtractionMemoryBlock The blogger says:VectorMemoryBlock The first step is to write history back in the chain. This design is natural because in the eyes of LlamaIndex, the history of dialogue, user facts, document clips can be processed into context.

It's also here: it's a knowledge-intensive aggent, business question and answer, document analysis, chat systems with search. It is not good at defining user memory, project memory, working board and permission boundaries for you. You can build these things with metada, session id, vctor stock, fact block, but the rules of governance still have to be fixed.

This is LlamaIndex: RAG / Context Framework grew out of Memoory. It is best organized by knowledge, history and extracting facts, rather than providing a complete system of memory for Argentina, which he cannot even free you to build.

OpenAI Agents SDK / Claude Agent SDK

OpenAI Ages SDK and Claude Age SDK are not suitable to continue with the five previous Memoory tables. They are more like exposing mature adults to developers: run loop, tool export, handoff or subagent, session, sandbox/workspace, rather than designing a set of memories from zero. Unlike AutoGen, SK, LlamaIndex, they don't give you a frame abstraction, then you spell an anent; they're a working anent, then they let developers around running tools, status and boundaries.

OpenAI, the center of this area is RunnerI'm sorry. Model to determine whether tool call, handoff, give final output; SDK execution tool, return the results back to the model, continue to circulate and process max_turnsThe project is being conducted in the country, and the projects are being carried out in the country. Claude Age as a better quality: wrapping down the tools, angent loop, context management, file operations and command execution of Claude Code as a library. They're not teaching developers to build antent, but to turn mature antent runtime into a programmable product.

So the short-term memory here is not the Memoory module, but the Transcript/session. OpenAI is available Session、conversation_id、previous_response_id、to_input_list() Continue the dialogue; Claude sided with session consultation, resume, fork. They address the continuity of short-term memory. Context control is also more like running hooks. OpenAI has session_input_callback、call_model_input_filter, handoff input filler, which determines how history is merged with new inputs and which elements are cut or rewritten before the model is called. Claude side relies more on Claude Code runtime management for contact management and session management. They give control points, not memory governance strategies; what is written in the context and which is left to local objects or external systems, is still to be judged by the developers.

Work memory was transferred to the workspace and file system. Normal OpenAI Agencies SDK does not have an inline whiteboard, and can be placed in local settings, tools, run state, external databases. Sandbox Ages has a separate sandbox file system: workspace, files, shell, and snapshots can all save the task site. Claude Code Agent SDK is naturally editing around file systems, commands and codes, and the task status is usually in repo, temporary files, command results and session translation. We're dispatching Claude Code, so his work memory is actually Claude Code.

Longer-term memory is more like workspace, and experience is sunk, not traditional database style. UserMemoryI'm sorry. OpenAI Sandbox Memory() The ones that will make the Prior run are the ones that will make it. MEMORY.md、memory_summary.md And rollout summaries, follow up run and read on demand. Claude Code. CLAUDE.md, project / user memory, auto memoory is similar: load the enduring context on startup and keep an angent less devious. External knowledge is the same, OpenAI's file reading/ search items for OpenAi's file search/ file readings for Claude are essentially the tools to get information into ant runtime. As for knowledge banks such as RAG, it requires user-specific configuration, and after all Runtme is self-directed to local tasks, not necessarily related to the knowledge base used for enterprise-level projects.

OpenAI AGENTS SDK and Claude Agent SDK: short-term to session / transition, work status to workspace / field system, workspace memory, workspace memory, workspace experience to work with, workspace experience to work with, work experience to work on, work to work on, work to work on, work to work on, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work to work, work, work to work, work to work, work, work to work, work to work, work to work, work, work to work, work on, work onCLAUDE.md- Yeah, auto memoory. They are more like mature Agent's programmable shell than the Memoory-first's Agent Framework. If you want to be rigorous with user memory, project memory, permission segregation, conflict consolidation and forgotten tactics, you still have to design yourself outside SDK.

Pydantic AI / PocketFlow

Pydantic AI does not need to be extended according to the five above categories, as it does not provide a complete memory management. It really offers a typed anent programming experience: message history that can be used to create a new technology that can be used to create a new environment. all_messages()、new_messages() and message_history In addition to multiple operations, the long-term memory, user memory, RAG, whiteboard, conflict consolidation, and permission segregation are all developed by the developers themselves to access databases, vector banks, searchers or external memory services. In other words, Pydantic AI does a beautiful job of using the history of news as structured data, not deciding what to remember, how to write, how to recall, how to delete.

PocketFlow is the same because it is the very simple process layout, not the Agent memoory process. It's all about Node, Flow, Action and... shared storeThe memory is at best a state of operation shared between nodes: intermediate results, return of tools, branch output, trace, end product can be put in. But... shared store It is not a long-term memory, nor is it a user image, much less a semantic search layer. To save between sessions, user isolation, RAG, memory extraction, expired deletion or conflict management, the developers must themselves take shared store Access to external databases, documents, vector banks or operating systems. The virtue of PocketFlow is transparency and lightness, and the disadvantage is that it barely manages your memory. But memories aren't always what you need.

Agent Memory and Runtime Atlas: External Memory, Runtime Systems, and Framework Memory

Agent Memoory and Runte technology inventory: hang memory, running time research and capacity-building in a framework

Emory Products and Research (in position)

Mem0 / OpenMemory

Zep / Graphiti

Letta / MemGPT

TencentDB Agent Memory

Cognee

Supermemory

MemOS

Generative Agents

Reflexion

MemoryBank

A-MEM

LongMemEval

HaluMem

MemoryAgentBench

MemConflict

LongMemEval-V2

Agent Runtime Research and Products (in position)

AIOS

SWE-agent

OpenHands

Agent S

OSWorld / BrowserGym

Magentic-One / Magentic-UI

Claude Code

Codex CLI

OpenClaw

Hermes Agent

Pi.dev

Agent Framework provides memory capability (in place)

LangGraph / LangMem

Mastra / Agno / AgentScope / CrewAI

AutoGen/Microsoft Agent Framework/Semantic Kernel

AutoGen

Semantic Kernel

Microsoft Agent Framework

Dify / Coze / n8n

Dify

Coze

n8n

LlamaIndex

OpenAI Agents SDK / Claude Agent SDK

Pydantic AI / PocketFlow

References

LangGraph / LangMem

Mastra / Agno / AgentScope / CrewAI

AutoGen / Semantic Kernel / Microsoft Agent Framework

Dify / Coze / n8n

LlamaIndex

OpenAI Agents SDK / Claude Agent SDK

Pydantic AI / PocketFlow