Comparing AI agent memory: Mem0, Letta, agentmemory, and Contextberg
These tools remember different things for different users. Compare the memory boundary, integration point, and best starting question before choosing one.
この記事を日本語で読む →
First ask whose memory you are building
“AI agent memory” can mean an application remembering a user, a stateful agent retaining context, a coding assistant keeping project knowledge, or a person recovering what happened across their desktop. Mem0, Letta, agentmemory, and Contextberg approach different parts of that problem. This comparison is based on their public documentation as reviewed in September 2026; integrations can change.
| Product | Main starting point | Memory boundary | Typical question |
|---|---|---|---|
| Mem0 | Add memory to an AI application | Application memories accessed through its APIs | “What should my assistant remember about this user?” |
| Letta | Build and run stateful agents | Agent state and memory blocks | “How should this agent maintain context across interactions?” |
| agentmemory | Add persistence to coding-agent workflows | Agent sessions, hooks, and related project context | “What did this coding agent learn from earlier work?” |
| Contextberg | Recover the person's desktop work across tools | Local Record, reports, and Memory | “What was I looking at and doing before I switched agents?” |
This is a comparison of *entry points*, not a feature score. Each product may support more integrations than the table can show.
Application memory: Mem0
Mem0 documents APIs for adding and retrieving memories in an AI application. It is a natural option when you control the assistant or product and want to decide which user information it should retain and recall. Evaluate its storage, retrieval, and privacy settings in the context of your application.
Stateful agents: Letta
Letta centers on agents with persistent state and memory blocks. Its documentation also covers coding and personal assistant uses, so it would be inaccurate to call it only an SDK for customer-facing apps. Start here if the agent itself, its context, and its runtime are what you want to design.
Coding-agent persistence: agentmemory
The agentmemory project focuses on retaining useful context from coding-agent workflows and exposing it through its supported integrations. Verify the current list of hooks and clients in its repository before choosing an installation path; those details change quickly.
Work memory: Contextberg
Contextberg starts from the person's work, including app and screen activity, browser history, typed input, and supported agent conversations when enabled. Record, Remember, Memory, and Chat use that local history. An MCP bridge can give compatible agents access to selected context. This is useful when the missing information is outside one agent's own session.
Choose by the question you need answered, then test a real workflow. For example, a coding-agent memory might recall a previous tool result, while a desktop work memory might also show the documentation and app screens you viewed before that agent session. Neither should replace checking current source files, tests, or documentation.