akitaonrails/ai-memory

AI coding assistants like Claude Code and Codex have no memory between sessions. When you close a session and open a new one, you have to re-explain the project, the decisions already made, and what was being worked on. ai-memory fixes that by maintaining a persistent wiki for each project, written in plain Markdown and stored in a git repository. Every prompt, tool call, and session boundary is captured automatically without any manual note-taking step. When a session ends, the relevant entries are compiled into coherent narrative pages in the wiki. When the next session starts, whichever AI tool you use next receives a handoff summary showing where things left off, what decisions were made, and what questions remain open. This works across different AI tools: you can stop mid-task in Claude Code and pick up in Codex without re-explaining anything, because both read from the same shared wiki. The server is written in Rust and runs locally via Docker. It exposes an MCP (Model Context Protocol) endpoint that supported AI clients connect to, and it provides a built-in read-only web interface for browsing the wiki in a browser. Full-text search over wiki pages is available both from the agent CLI and from the web UI. No vector database is required, search runs on SQLite with FTS5. Each project's wiki is stored under a path keyed by stable identifiers, so two projects with identically named files never collide. Wiping a project removes only that project's data. The server can also run on a home server or another machine on the network so multiple devices share one memory store, with an optional bearer token for access control. Installation involves running the server as a Docker container and then running two commands that wire the MCP connection and lifecycle hooks into whichever agent CLI you use. The hooks fire automatically on session start and end, so there is nothing extra to do during normal use. The wiki files are plain Markdown, which means they can be opened in any editor, searched with standard tools, and backed up like any other directory. The full README is longer than what was shown.