When multiple agents share one database, pass 2 loads every other agent's active claim on the same entity and labels it equivalent, compatible, or contradicts. Compatible claims stay active together. Equivalent claims write nothing. Strategies below run only on claims labeled contradicts.

Strategies

Configure per entity with entityStrategies. Unconfigured entities default to recency.

Recency (default)

Newest contradicting claim wins. Those claims are superseded; incoming becomes active. Compatible claims on the same entity are left alone. No disputes row.

Trade-off: a newer casual-chat message can overwrite an older authoritative fact.

Authority

Compare agent_authority weights for the writing agent vs each conflicting claim's sourceAgent on that entity.

• Incoming weight strictly greater than every contradicting agent → REPLACE
• Otherwise → DISPUTE

Authority must be seeded manually — default weight is 0.

import { seedAuthority, EXAMPLE_AUTHORITY_SEED } from "vault-memory";

await seedAuthority(db, EXAMPLE_AUTHORITY_SEED);
// or custom:
await seedAuthority(db, [
  { entity: "user.timezone", agentId: "agent-a", weight: 0.9 },
  { entity: "user.timezone", agentId: "agent-b", weight: 0.2 },
]);

const memory = new Memory("agent-a", {
  db,
  vectorStore,
  apiKey: process.env.OPENAI_API_KEY,
  entityStrategies: { "user.timezone": "authority" },
});

Classic demo

Disputes

When authority cannot pick a winner, Vault:

1. Stores the incoming claim as disputed (not indexed)
2. Leaves the incumbent active and searchable
3. Inserts a disputes row with resolved_at null

Resolve at the app layer — see Dispute resolution. Vault does not auto-judge with an LLM.

FAISS in multi-process setups

Each process holds its own in-memory FAISS index. Call vectorStore.rebuild(db) before reads/writes when another process may have written claims.