MemoryOps AI
Governed memory runtime for AI assistants — policy-before-storage, deletion proof, tenant isolation, and leakage evals
MemoryOps AI is an open-source governed memory runtime for production AI assistants. It controls what becomes memory, what enters context, what must be forgotten, what influenced an answer, and what evidence proves each decision — treating memory as governed state, not just a vector database.
Enforced Invariants
- Tenant isolation — One user’s memory is never returned to another user or tenant.
- Deletion guarantee — Deleted memories are never retrieved again, with tombstone lineage as proof.
- Policy-before-storage — Unsafe or secret-like content is filtered before it is stored.
- Auditability — Every lifecycle mutation and its audit event commit together in one transaction, as an append-only, tamper-evident chain.
- Explainability — A memory usage trace shows which memories affected a response.
- Graceful degradation — A retrieval failure never blocks a response.
Results
| Signal | Result |
|---|---|
| Governance benchmark | 50/50 (100%) — critical suites (deletion/leakage, tenant isolation) perfect |
| Extraction quality (gemini-2.5-flash, 25-turn labeled set) | 0.94 precision / 0.94 recall / 0.94 F1 |
| SDK | memoryops-sdk on PyPI with a stable 1.x API contract |
| Deployment | Live playground on Railway (web · API · worker · Postgres · Redis) |
Postgres/pgvector and in-memory backends are fully tested in CI; Qdrant, LanceDB, and Weaviate adapters are contract-tested.
Tech Stack
Python · FastAPI · PostgreSQL + pgvector · Redis · Qdrant · LanceDB · Weaviate · JWT/JWKS · Prometheus · Next.js · Docker · Railway