Switch AI agents without losing your place

What it does ai-memory is a Rust service that sits between you and your AI coding agents. It hooks into the lifecycle of Claude Code, OpenAI Codex, Cursor, and a dozen other clients, automatically logging every prompt, tool call, and decision to a plain-markdown wiki backed by git. When a session ends, it compiles the noise into coherent narrative pages; when a new agent starts in the same project, it prepends a “where you left off” handoff so you don’t re-explain the architecture or repeat failed approaches.

The interesting bit Instead of a vector database or manual note-taking, the wiki is just markdown files in a git repository—grep-able, sync-able with rsync, and browsable in Obsidian or its built-in /web UI. The cross-agent handoff is the real trick: the project explicitly treats Claude, Codex, Grok, and others as interchangeable workers on the same shared memory, even when they refuse to read each other’s stdout. Grok and Zero currently need an explicit MCP call to accept handoffs, since they ignore SessionStart injection.

Key highlights

• Supports Claude Code, Codex, Cursor, Devin, Gemini CLI, OpenCode, and others via MCP configs and lifecycle hooks
• Stores memory as plain markdown in git—no vector DB, no write_note ceremony
• Per-project isolation with UUID-based paths; global scope for durable personal rules that follow you across projects
• Optional LLM-powered consolidation and auto-improvement scheduler that proposes wiki edits from finished sessions
• Built-in web UI with FTS5 search; runs locally or on a shared homelab/server with bearer-token auth

Caveats

• Claude Desktop and VS Code Copilot only get MCP access, not lifecycle hooks, so they can’t auto-capture or auto-inject handoffs
• Grok and Zero capture events but ignore SessionStart stdout, forcing handoff recovery through an explicit MCP call
• Native Windows support is experimental; WSL2 is the recommended path there

Verdict Worth a look if you routinely switch between Claude Code and Codex mid-project, or if you want a grep-able, git-versioned audit trail of why your agent made specific decisions. Skip it if you only ever use a single agent in short bursts and don’t mind repeating yourself.