Shared Mind
The retrieval layer every other agent shares — one second brain across your team and your bots
Shared Mind is the cross-agent memory and retrieval surface for the Hannah Imagineer ecosystem. It runs on its own Supabase pgvector instance (isolated from the operational DB), ingests documents from the customer's sources, and exposes typed retrieval to every other agent (Composer, Hermes, OpenClaw, Concierge) via a single RPC.
Overview
Shared Mind is what stops every agent from re-discovering the same context. A single ingestion run populates the second brain; every downstream agent reads from it. Vector store is isolated on its own Supabase instance for workload separation — embedding spikes never starve the transactional path.
Primitives
Cross-agent retrieval via a single typed RPC
Isolated pgvector instance — no contention with transactional DB
Source-attributed citations on every retrieval
Per-tenant namespaces with row-level isolation
Incremental ingestion with content-hash dedup
Hybrid keyword + vector search
Outcomes
- 01One memory layer for every agent — no per-agent re-ingestion
- 02Embedding workload isolated from the transactional path
- 03Every agent answer carries an auditable citation trail
Integrations
Autonomy & guardrails
- Cross-tenant retrieval
- Ingestion of sources outside the configured allow-list
Guardrails & requirements
Guardrails
- Per-tenant namespace isolation enforced at the query layer
- Source-attributed retrieval — no hallucinated citations
- Embedding model and chunking strategy versioned per namespace
Requirements
- Supabase project with pgvector extension
- Source connectors enabled for the customer's document surface
Technical specifications
Runtime
- Harness
- Hannah KB Vector Store
- Deployment
- Managed (Hannah KB Supabase) · BYO Supabase project supported
- Data residency
- Customer-controlled — vector store lives in the customer's Supabase region
- License
- Commercial · Hannah Imagineer EULA
- Version
- v1.0
Models & tooling
- Models
- text-embedding-3-largeclaude-sonnet-4-6 (reranker)
- Tooling
- pgvectorarchon-ingestrerank-rpccitation-trail
Reliability & limits
- Latency SLA
- ~150ms p50 retrieve · ~600ms p50 with reranker
- Rate limits
- 300 retrievals/min per namespace
Security & compliance
- Auth model
- Service-role RPC from Composer · row-level isolation per namespace
- Compliance
- SOC 2 Type II (in process)GDPR
Evals & benchmarks
Public benchmark numbers (BEIR / MTEB-style) for the Shared Mind pipeline are scheduled for the next release alongside Composer-native benchmarks. The internal eval set already covers retrieval recall@10 and answer faithfulness with LLM-as-judge against ground-truth citations; results will be published once the public eval methodology document lands.
Architecture notes
Memory architecture
Single-tenant pgvector namespace per customer in the dedicated Hannah KB Supabase instance. Tables: archon_sources, archon_crawled_pages with 1536-dim embeddings. The transactional Hannah DB is a separate Supabase project to isolate embedding workload from app reads.
Context strategy
Hybrid retrieval: keyword filter narrows the candidate set, vector similarity ranks within it, optional Claude reranker re-orders the top-K. The Composer-side context bundle is shaped per-subagent so each agent only sees the slice it needs.
Evaluation
Internal eval set covers retrieval recall@10 and answer faithfulness (LLM-as-judge against ground-truth citations). Per-namespace drift monitored against the published baseline.