Expand description
The one loader: build scorable Memory items from a MemoryWhale SQLite DB.
Both surfaces call this — the desktop app (documents + command_runs +
agent_turns) and the CLI (sessions + command_runs + bookmarks). The
two databases have different tables, so every source is queried
independently and a missing table is treated as zero rows (not an error).
That keeps a single loader honest across both schemas instead of two copies
drifting apart.
Ids are namespaced per source so explain(id) stays stable and unique
across sources; decode_id recovers the source + original row id for
display (e.g. mw replay <id> / mw show <id>).
Enums§
- Source
- Which store a memory came from, plus its original per-table row id.
Functions§
- decode_
id - Recover the source and original row id from a namespaced memory id.
- load_
memories - Load everything MemoryWhale remembers as scorable memories, tolerant of any source table being absent (so it serves both the desktop and CLI schemas).