Claudeの記憶を「メモ帳・日記・本棚」の3層に分ける — dejavu というOSSを作りましたdejavu is an open-source MCP server that organizes Claude's memory into three…

dejavu is a newly published open-source Model Context Protocol (MCP) server that gives Anthropic's Claude a structured, persistent memory divided into three distinct layers. The project matters because large language models remain fundamentally stateless between conversations, and their context windows, though growing, are finite. Tools that let an assistant remember earlier work, stated preferences, and accumulated knowledge address one of the most common frustrations for people who use Claude for ongoing projects rather than one-off queries.

The core idea behind dejavu is a memory hierarchy that its author describes with everyday metaphors: a notepad, a diary, and a bookshelf. The notepad represents short-term storage for the immediate working context—transient notes relevant to the current task. The diary sits in the middle as medium-term memory, capturing session-by-session records that give continuity across days or weeks. The bookshelf serves as long-term storage for durable knowledge and reference material that should persist indefinitely. Separating memory this way is intended to mirror how people naturally distinguish between fleeting reminders, journaled experiences, and consolidated knowledge.

This layered design echoes established concepts from both computer science and cognitive psychology. It resembles the memory hierarchy in operating systems, where fast, small caches sit above slower, larger stores, and it parallels psychological models that separate short-term, working, and long-term memory. The practical benefit is that an assistant can decide what to keep close at hand versus what to archive, rather than treating all remembered information as an undifferentiated pool. This can help manage the trade-off between retrieval relevance and context-window cost.

Understanding dejavu requires a little background on MCP itself. The Model Context Protocol is an open standard that Anthropic introduced in late 2024 to standardize how AI applications connect to external tools, data sources, and services. An MCP server exposes capabilities—such as reading files, querying a database, or, in this case, storing and recalling memories—that an MCP client like Claude Desktop or other compatible applications can call. Because the protocol is open, developers can build servers in various languages and share them, and a growing ecosystem of community servers now covers everything from GitHub access to web search. dejavu fits this pattern as a specialized memory server.

In use, a memory-oriented MCP server typically works by letting the model write entries during a conversation and retrieve them later, so that context carried over from a previous session does not have to be pasted in manually. By assigning entries to the appropriate layer, dejavu appears designed to let Claude surface the most relevant material without overloading the prompt with everything it has ever recorded. The exact storage mechanism, retrieval strategy, and whether it relies on vector embeddings or simpler keyword lookup would depend on the implementation details in the repository.

dejavu enters a crowded and fast-moving space. Anthropic has been adding native memory features to Claude, and OpenAI offers memory in ChatGPT, while dedicated projects such as Mem0, Letta (formerly MemGPT), and Zep tackle persistent agent memory with vector stores and structured recall. Many of these systems similarly borrow the idea of tiered or hierarchical memory. An open-source, MCP-native option is notable because it can run locally, keeps data under the user's control, and integrates with any MCP-compatible client rather than being tied to a single vendor's product.

Prospective users should weigh a few practical considerations. Persistent memory raises questions about where data is stored, how it is secured, and how outdated or incorrect entries are pruned, since a memory that accumulates errors can degrade rather than improve an assistant's usefulness. As with most early-stage open-source tools, the maturity, documentation, and long-term maintenance of the project are worth checking before relying on it for important workflows.

For developers experimenting with Claude as a long-running collaborator, dejavu offers a conceptually clean approach to a genuine problem. Whether the three-layer metaphor proves more effective in practice than flat or purely vector-based memory will likely become clearer as users test it across real projects. Even so, the framing is an accessible entry point into the broader challenge of giving AI assistants a reliable sense of continuity, and it reflects a wider trend toward treating memory as a first-class component of practical AI systems rather than an afterthought.