Memory architecture
Memory provides project-scoped committed code and Markdown discovery backed by local SQLite and FTS. Parsing is deterministic and uses no model or remote API. Successful snapshots survive failed refreshes, and indexing attempts are recorded independently. Authored knowledge lives in implementation-local Markdown, not agent-written database rows. AI tools and natural-language discovery are query-only; indexing and automation management use explicit app commands/settings and commit hooks.
Entry points
init.ts— Plugin lifecycle, core context, and command/AI registration.adapter.ts— CLI parsing and per-command dispatch.definition.ts— Aggregated command definitions.ai.ts— Read-only tool schemas/dispatch, project-status context, and skill definitions.db.ts— Local SQLite opening and derived-index table creation.discovery.ts— Independent code/Markdown parsing attempts and joint ranked search.search-provider.ts— Core universal-search adapter over existing snapshots and read-only code/Markdown inspection views; owns workspace-scoped resource IDs.commands/— Help, natural-language discovery, code, Markdown, and project-status surfaces.commands/shared/— Natural-language argument normalization.
Modules
- Code intelligence owns TypeScript graphs, project status, external-target navigation, and explicit opt-in Git hooks.
- Markdown intelligence owns committed Markdown snapshots, headings/chunks, local search, and document links/backlinks.
See Memory design and progress and user guide.