Data model

Waymark stores everything in one local SQLite database ($WAYMARK_HOME/waymark.sqlite3). All timestamps are UTC ISO strings. The schema is intentionally simple and relational; there is no hidden vector database.

Tables

Table	Purpose
entries	Memory cards: title, raw_text, summary, type, mood, source, timestamps.
tags	Unique tag names.
entry_tags	Many-to-many link between entries and tags.
decisions	First-class decisions: context, options, status, final choice, confidence, review date, outcome.
decision_entries	Links decisions to the memory entries that informed them.
reflections	Saved reflections per period, with wins, patterns, and suggestions.
sources	Imported files: type, path, original filename, metadata.
entry_embeddings	Explicitly generated vectors per entry and model.
entries_fts	FTS5 full-text index over entries — derived, rebuilt by triggers.

How pieces connect

• A memory entry can carry a source of manual, a local-AI draft, or source:<id> pointing at a row in sources (for imported files). Retrieval resolves that into a label such as pdf: brief.pdf.
• Tags are normalized (lowercased, de-duplicated, sorted) and shared across entries through entry_tags.
• Decisions link to the entries that informed them via decision_entries, keeping each decision's evidence trail visible.
• Embeddings are opt-in. Nothing is embedded until you run waymark embeddings backfill --apply.

Full-text search

entries_fts is a contentless FTS5 index kept in sync with entries by insert, update, and delete triggers. Because it is derived, it is excluded from backups; restoring entries re-fires the triggers and rebuilds the index automatically.

See the canonical schema in src/waymark/storage.py.