A persistent recall layer designed to give every Arcanada agent memory across sessions. We benchmarked three candidates (Hindsight, Graphiti, Mem0) and selected Hindsight, then validated the Model Connector path: Claude Haiku is stable, the Cursor connector turned out unsuitable for structured output. Production deployment is pending — research foundation is solid, hardening is the next horizon.
- Status
- Research; production deployment pending
- Autonomy (AAL)
- L1 → L3
- Backend
- Hindsight
- Candidates compared
- Hindsight, Graphiti, Mem0
01
Memory across sessions
The goal is to give every Arcanada agent recall between sessions, with semantic retrieval and per-agent isolation.
02
Forgetting by policy
TTL plus relevance decay are planned to decide what fades.
03
Structured output is required
The Model Connector path is validated: Claude Haiku is stable, while the Cursor connector turned out unsuitable for structured output.
How the backend was chosen
Three candidates (Hindsight, Graphiti, Mem0) were benchmarked, and Hindsight won. The Model Connector path was then validated end to end.
Production deployment is still pending: the research foundation is solid, and hardening is the next horizon.
Capabilities
Planned
- Cross-session memory retention per agent
- Semantic memory retrieval via embeddings
- Forgetting policies — TTL plus relevance decay
- Per-agent isolation, no cross-tenant bleed
Status not stated
- Hindsight backend (winner of internal benchmark)
- Tested integration path with Model Connector + OpenRouter
- Structured-output requirement documented (no Cursor/Claude CLI)
Current autonomy level
Weakest link
Research phase — no production hardening yet. A known race condition in shared workspaces during parallel commits is something any production memory layer must defend against.
Roadmap to L3
- L1 → L2 — production deploy on the ecosystem's managed database host, /health endpoint, basic deploy reporting.
- L2 → L3 — full observability, schema-validated reads/writes, fallback to no-memory mode on outage so agents degrade gracefully instead of crashing.
- Verification gate — soak test: 1000 concurrent agent sessions for 72 h with zero memory cross-contamination.
Next step
Review the current build and its development plan
BUILDING means active development, not a promise of public availability. Use the current capabilities and roadmap above as the reference.
Review the path to the target level