Skip to content

Layer cake

Talos layer cake — thin clients, talosd (channel adapters / runtime / KeeperHub middleware), PGLite + Drizzle, wallet + knowledge cron, hot MCP tool servers

#DecisionChoice
1Agent loop substrateVercel AI SDK v6 (ai ≥ 6.0, @ai-sdk/mcp for MCP client)
2PersistencePGLite single store (knowledge + conversations)
3ORMDrizzle (schema-as-TS, auto-generated migrations)
4Embedding modelOpenAI text-embedding-3-small (1536-dim)
5RetrievalHybrid: pgvector (HNSW) + tsvector (GIN)
6Cross-thread recallAdaptive: cosine ≥ 0.78, top-3, always-search
7Summarization triggerTurn count: every 20 runs
8LLM call stylestreamText everywhere
9KeeperHub defaultAudit-by-default — undeclared tools routed

For the full design — schemas, event taxonomy, retrieval strategy, daemon lifecycle, audit-by-default contract — see docs/architecture.md in the repo.