Architecture
UnifiedMemory is an agent-native memory fabric, not a raw vector database. The hot path runs at the edge; heavier cognition and Courtroom work stay in the origin backend.
Runtime responsibilities
| Layer | Responsibility |
|---|---|
| Cloudflare Worker | Fast auth, scoped hot reads, MCP/edge API routing, D1/Vectorize access, lightweight runtime status. |
| D1 | Canonical production metadata: keys, sessions, memory catalog, review/quarantine state, trust and lifecycle fields. |
| Vectorize | Semantic candidate retrieval using coarse metadata filters. D1 hydration remains the final authority. |
| R2 and Queues | Evidence snapshots, proof artifacts, async fanout, delayed lifecycle work, and retry/dead-letter surfaces. |
| FastAPI origin | Heavy cognition, compatibility APIs, Courtroom reports, Google Memory Bank bridge, migrations, proof tools. |
| MiniMax | Default UnifiedMemory-owned cognition and AI judge lane. It can recommend, downgrade, and explain, but not bypass guardrails. |
| Google Memory Bank | Managed memory co-processor for session ingestion, generated memories, profiles, and revisions. It is evidence, not automatic truth. |
Safety boundaries
- Effective identity comes from authenticated key/session metadata.
- Agent-private working memory is not shared by default.
- Test, smoke, benchmark, diag, eval, forensic, and system-generated rows are excluded from assistant-facing recall.
- Courtroom deterministic guardrails run before AI judgment.
- Google and MiniMax disagreement creates review/conflict evidence rather than automatic promotion.
Next steps
- Read the Memory model.
- Read Cross-agent isolation.
- Connect an agent with the agent guide.