Packages and data flow

Package boundaries

platform/core owns shared types, SQLite access, migrations, search, and identity primitives. platform/daemon owns the HTTP API, background runtime, authorization, and writes. surfaces contains the CLI, React dashboard, desktop application, and other human-facing clients. integrations contains harness adapters; libs contains reusable packages; dist contains shipping artifacts; web contains the public sites; and memorybench is a development benchmark harness.

The dashboard, CLI, and integrations are clients of the daemon. They should not implement competing memory or authorization behavior.

Current data flow

harness, CLI, dashboard, or connector
  -> daemon HTTP boundary
  -> agent-scoped evidence and SQLite state
  -> FTS, embeddings, source/document indexes, and optional hints
  -> bounded recall with authorization
  -> agent or user receives evidence-backed context

completed episodic evidence
  -> Dreaming selection
  -> audited ontology operations
  -> current scoped ontology and graph traversal inputs

Evidence capture and semantic maintenance are separate. A saved memory or source artifact remains attributable evidence. Dreaming is the only automatic semantic writer. There is no active extraction worker, decision worker, or automatic relation writer in the runtime.

Workspace state

SQLite rows are canonical application state. Search indexes, vectors, caches, generated projections, and dashboard payloads are derived and rebuildable. User-facing identity files, skills, source artifacts, imports, logs, and backups remain workspace artifacts rather than database replacements.

See Pipeline and storage for worker and persistence boundaries.