Introduction
Agent-attributed memory for AI agents with semantic search and multi-agent conflict resolution.
Vault is an open-source memory layer for AI agents. It extracts structured facts from conversation, stores them in your database, and retrieves them with semantic search — with first-class support for multiple agents writing to the same store.
Most agent memory is append-only: new facts pile on, contradictions get duplicated, and last-write-wins hides who said what. Vault treats memory as claims with attribution, lifecycle, and conflict policy.
Why Vault
Structured facts
Extraction returns entity, text, and type (semantic or episodic) — not raw chunks.
Two-pass classify
Reconcile within an agent first (ADD / UPDATE / REPLACE / NOOP), then label other agents' facts.
Multi-agent policy
Compatible facts stay. Contradictions use recency or authority — disputes when trust cannot pick a winner.
Your database
Drizzle + SQLite or Postgres. FAISS vector index locally; SQL is source of truth.
When to use Vault
Vault fits when you have:
- Multiple agents (scheduling, chat, tools) sharing one memory store
- Facts that can contradict or refine over time
- A need for audit trails (
superseded,disputed) instead of silent overwrites
For a single agent with simple append-only notes, a lighter store may be enough. Vault shines when accuracy and attribution matter.