Interactive visualization of SPINE's 7-tier memory architecture with unified facade and verdict routing.
MemoryFacade.search() fans out a single query across all 7 tiers, normalizes scores, and returns ranked results.
After task evaluation, the VerdictRouter directs results to the appropriate memory tier based on the verdict.
| Tier | Name | Backend | Scope | Persistence |
|---|---|---|---|---|
| 1 | KVStore | SQLite / File | Namespace-scoped | Durable |
| 2 | Scratchpad | In-memory list | Task-scoped | Volatile |
| 3 | EphemeralMemory | In-memory + decay | Session-scoped | TTL Decay |
| 4 | VectorStore | LanceDB / Keyword | Collection-scoped | Durable |
| 5 | EpisodicMemory | SQLite + FTS5 | Goal-scoped | Durable |
| 6 | DeepMemoryStore | PostgreSQL + pgvector | Project-scoped | Durable |
| 7 | GraphMemory | PostgreSQL + NetworkX | Cross-entity | Durable |
Unified search across all 7 tiers. Score normalization and tier weighting merge results from different backends into a single ranked list.
facade.search("query", top_k=10)
Read-only cross-project memory federation. Parallel fan-out to remote Minna MCP servers with budget controls and provenance metadata.
fed.search("query", limit=10)
Routes evaluator verdicts to the appropriate tier. Accept stores episodes, Revise updates scratchpad, Reject logs to ephemeral memory.
router.route(verdict, task, result)
SPINE's memory architecture uses 7 tiers to match different temporal scopes and access patterns. Fast lookups use Tier 1 (KVStore), session-scoped observations use Tier 3 (Ephemeral), and long-term semantic knowledge lives in Tier 6 (DeepMemoryStore) with PostgreSQL + pgvector.
The MemoryFacade provides a single search interface across all tiers. The VerdictRouter closes the feedback loop by directing task outcomes to the appropriate tier based on the evaluator's verdict. FederatedMemory extends reach across project boundaries via parallel MCP server queries.