Agent Memory Vault

Agent Memory Vault: Shared Memory for Claude Code and Codex

English | 简体中文

Agent Memory Vault is a local-first, Git-backed long-term memory system that Claude Code and Codex can safely share. Markdown remains the source of truth; SQLite provides structured and full-text retrieval; optional EmbeddingGemma + Zvec adds local semantic search.

The repository contains only reusable templates, scripts, and fictional examples. Your real memories, paths, credentials, project names, and conversation content stay in your private local vault.

Why it exists

AI coding agents are useful inside one session, but durable collaboration needs more than chat history. This project provides a verifiable memory lifecycle:

- Start important work by retrieving only the relevant long-term context.
- Preserve stable facts, decisions, workflows, project state, and agent lessons as readable Markdown.
- Let Claude Code and Codex use one vault, one Git history, and one retrieval index.
- Prevent sessions from accidentally committing each other's changes with session-scoped claims.
- Validate high-impact writes with source checks, content-bound intents, approvals, and immutable receipts.
- Keep every derived store rebuildable from Markdown.

Obsidian is optional. The vault is an ordinary Markdown directory and works with any editor.

Design

The repository is organized around a small set of auditable components:

See Architecture, Privacy, and Automation for the detailed model.

Quick start

Requirements: Python 3.10+ and Git.

The recommended macOS/Linux entry point is read-only planning followed by one
explicit apply. Choose host hooks explicitly; this example intentionally uses
no host hooks:

--plan always runs the migrator from the source checkout with an existing
Python 3.10+ (use --python…