Stimulus
Joshua Gunn's Leviathan (Apache-2.0) is a recall stack built around searchable memory cards. We ran a read-only bench on moya against Athena's real corpora — not a toy dataset — to see whether Leviathan-style FTS would surface answers Athena's built-in search misses.
The pain
Athena's conversation_search often returns empty while the answers sit in message content JSON. Tool samples show roughly 65% empty returns on conversation_search. The root cause is not missing data: the tool queries an empty text column while the substance lives elsewhere in the message payload.
Lab result
Leviathan-style FTS cards recovered the false-negative set:
- hit@5: 18/19 (95%)
- empty returns: 0/19
- Only miss: B13 (deep pre-migration wording)
- Skipped: B08 (never migrated from ChatGPT)
Corpus under test: 30,664 Letta messages, 282 archival rows, 165,353 origin messages.
What this is not
Leviathan is SQLite + FTS5 cards. Postgres is an ingest pipe in that design — not a reason to abandon Postgres for markdown just to fix search. The fix is recall that actually indexes the content Athena already has.
Charter amendment
Recall-first. Substrate migration is reopened and must not block FTS. Search that works beats a cleaner store that still cannot answer.
Next program
- Sanctum-owned Python Leviathan-class recall with SQLite and Postgres backends.
- Slim Sanctum Letta — telemetry-driven cut to a kernel plus microservices.
Ops
The moya lab was torn down the same day (~788MB). Athena was untouched.