mcncarl/agent-memory-vault

What it solves

AI coding agents often lack durable, shared long-term memory across different sessions or different tools. This project provides a local-first system that allows agents like Claude Code and Codex to share a single, verifiable memory vault, preventing the loss of project state, decisions, and lessons learned between sessions.

How it works

The system uses Markdown files as the primary source of truth, which are stored in a private Git-backed directory. To enable efficient retrieval, it maintains a SQLite database for structured and full-text search, with an optional local semantic search layer using EmbeddingGemma and Zvec.

To ensure data integrity and prevent agents from overwriting each other's work, it implements a "claim" system where agents must claim files before they can be committed during a "closeout" process. It also features a two-phase write workflow (read-target $\rightarrow$ prepare $\rightarrow$ apply) to provide a safety boundary against accidental misuse.

Who it’s for

Developers using AI coding agents (specifically Claude Code and Codex) who want a persistent, auditable, and private long-term memory system for their AI assistants.

Highlights

• Local-first & Private: Memories are stored in a private local vault, separate from the public code repository.
• Git-backed: Uses Git for version control, history, and rollbacks of memory files.
• Hybrid Search: Combines keyword search (SQLite/FTS) with optional local vector-based semantic search.
• Session-scoped Claims: Prevents session conflicts by ensuring only claimed files are processed during closeout.
• Verifiable Writes: Includes source checks, content-bound intents, and immutable receipts for high-impact changes.