Substrate: architecture
Shared by Pavel · Updated 2026-11-12
Substrate’s vendor material names Gmail, Notion, WhatsApp, uploads, and consented agent sessions as inputs; it gives a 50 MB per-file upload limit.
Capture, distill, correct, store, retrieve
- Capture is consent-based, and each source is kept exactly as it arrived. Gmail and Notion are read-only; agent sessions are captured only with consent.
- Raw messages, documents, and events are retained alongside derived facts. Distillation breaks material into atoms—people, organizations, projects, tasks, and decisions—with evidence, a date, and confidence.
- A human edits a readable wiki page; that edit becomes a durable fact and feeds back into the graph. Deleting a sentence is treated as a correction, rather than something the system recreates later.
- Storage is described as three linked layers: raw source material, a graph connecting entities and decisions, and readable wiki pages.
- Agents receive relevant context at session start and can pull more on demand. The vendor-facing description says agents read pinned material and facts relevant to the current message, with context returned each turn; they can open an atom’s source excerpt.
- Substrate’s own description of retrieval is dense, lexical, graph, and temporal retrieval fused and reranked, using local embeddings. The FAQ says models and weights are not published.
Access and sharing
- Agents connect through MCP; each connection is approved in a browser with OAuth and can be revoked. Agent keys are stored hashed.
- Each account is a tenant; Postgres row-level security is enforced by the database, and the application role cannot bypass it.
- A page and its descendants can be shared as a live or frozen branch, with or without raw files, and sharing can be revoked.
Operational behavior and stated limits
- Model output is checked before storage; the model does not select record identifiers or decide which account data a record belongs to.
- Jobs persist across restarts; retries are bounded, and failed work has a visible Retry action. Pages show queued, generating, ready, failed, or interrupted status.
- During source re-ingestion, the last good version remains active until its replacement completes; a crash before commit leaves the prior successful version active.
- Search pages and cards are derived views that can be rebuilt from underlying facts; replacement views are built before readers switch over.
- The beta runs in one region without a high-availability replica; no SLA, uptime figure, recovery-time commitment, or multi-region failover is claimed.