mem0 + Vault
Published governance lift
The provider and candidate pool stay fixed. Vault blocks stale and superseded results, then selects the next valid candidate.
Memory engine integrations
Keep the engine. Add a trust layer.
Vault does not need to replace retrieval. It can govern what mem0, AgentMemory, Letta / MemGPT, vector RAG, or a custom engine returns before an agent treats it as usable memory.
Shared augmentation pattern
One governance contract across changing engines.
The provider ranks candidates. Vault evaluates canonical state and returns allowed evidence, block reasons, provenance, and a rollback path.
A · Provider
Retrieve candidates
Native extraction, embeddings, ranking, and storage stay with the selected engine.
→
B · Vault
Apply trust policy
- Approval and lifecycle
- Access and sensitivity
- Freshness, TTL, supersession
- Privacy and deletion
- Provenance and audit
→
Agent output
Use governed evidence
Permitted memory, citations, explicit exclusions, and recoverable decisions.
Evidence status
Implemented is not the same as published.
Status reflects the current checked-in artifact chain. Missing evidence remains visible.
| Engine | Native role | Vault augmentation | Evidence state |
|---|---|---|---|
| mem0 2.0.12 | Controlled retrieval with native embeddings and ranking | Post-retrieval governance guard | Published Five clean-state repeats; 4/6 → 6/6 Valid Recall@1; forbidden exposure 2/6 → 0/6. |
| AgentMemory 0.9.27 | Global smart-search retrieval | Guard and optional Vault retrieval fusion | Diagnostic Five developer runs exist, but they predate blinded-input and clean-source gates; the pinned dependency audit also reports 1 critical and 6 high vulnerabilities. |
| Letta / MemGPT 0.16.8 | Stateful agent runtime with Archive / Passages | Guard over Archive candidates | Unmeasured Adapter exists; a pinned PostgreSQL / pgvector server and lifecycle proof are still required. |
| Custom RAG | Any engine emitting the neutral run-artifact schema | Guard-only or reciprocal-rank fusion | Provider-specific Publication requires the same blind input, isolation, source binding, and repeat gates. |
AgentMemory + Vault
Adapter works; claim is gated
The adapter maps provider observation IDs back to neutral source IDs and fails closed on ambiguous results. A clean blinded rerun and a dependency tree with zero critical or high advisories are required.
Letta + Vault
Runtime and governance can stay separate
Letta can own stateful-agent execution while Vault supplies a portable trust contract. No quality claim is made before a live pinned run.
What the integration changes
Retrieval relevance becomes governed usability.
The value is not “Vault makes every engine rank better.” The defensible claim is narrower: Vault can prevent an engine's relevant-looking but invalid memory from becoming agent evidence, while preserving inspectable reasons and canonical state.
Bring another memory engine.
Emit the neutral artifact contract, run A versus A + Vault, and publish only after the gates pass.