Anamnesis — AI skill for Claude Code

Anamnesis

[](https://github.com/berketpbs/anamnesis/actions/workflows/ci.yml)

**Shared memory for AI coding agents.**

Every agent session starts from nothing. The decision you settled yesterday, the approach that already failed, the reason a file looks the way it does: all of it is gone when the terminal closes, and gone again when you switch from Claude Code to Codex. Anamnesis keeps that memory in one place that every agent on the project reads from and writes to. It is a plain markdown wiki under git, and it lives on your machine.

• One memory, every agent. Claude Code, Codex CLI, Gemini CLI, Cursor and OpenCode record into the same project memory, and whichever starts next is handed what the last one did.
• It comes to the agent. A new session starts with a handoff and the project's recorded decisions. Each prompt is answered with the pages memory already has on it, and with nothing when memory has nothing to say.
• You can read it. Memory is markdown in a git repository: browse it, edit it, diff it, revert it. Nothing is hidden in an embedding store.
• Local and private by default. Everything is redacted before it is stored, the server listens on 127.0.0.1, and no model or account is required.

How it works

  agent hooks ──▶ capture ──▶ raw/ transcripts  (redacted, append-only)
                                  │
                   session ends   ▼
                            consolidation ──▶ wiki/  (markdown + git)
                            (count, or a model)   │
                                                  ▼
  next session ◀── handoff at start · recall at each prompt · MCP tools on demand

1. Capture. Hooks in each agent send lifecycle events (prompts, tool calls, their results, session start and end) to a small local server. Each event is redacted and appended to a transcript that is never rewritten.
2. **Consol