Throughline's recall receipts separate an empty archive from a search that never ran

A recall function that returns a list uses the same value for "no rows matched" and for a search whose embedding call timed out and was caught by a handler. The agent reading that list sees zero rows either way. According to Throughline's author, most memory stores return both cases as the same empty list ^[2].

Throughline puts coverage into the return value. Each receipt records the retrieval path that actually ran, how many candidates it examined, and what it excluded under which rule ^[4]. The verdict travels with the rows to the boundary where results become prose, and the guard there rejects an UNKNOWN reported as "nothing found" ^[5]. The memory types are tuned for outages as well. A rejected hypothesis such as "Restarting the pods did not help" keeps its value for a year, while an entity fact like which host is primary has a 14-day half-life ^[3].

The same collapse turned up in CockroachDB's managed MCP server. The author wrote a client for it as a second channel that has to agree with the direct connection, though it never reached the running demo ^[7]. With failures arriving as HTTP 200 and no result ^[8], a client written the obvious way, `rows ?? []`, turns an outage into "that row doesn't exist" ^[9]. Status codes, on this evidence, are decorative. The server's `select_query` also adds `LIMIT 25` when the caller gives none ^[10]. A caller that gets exactly 25 rows back cannot tell a complete answer from a capped one, and in Throughline's terms I'd call that PARTIAL.

The index work is careful. Filtered recall had been planning as a full scan, and the docs do say filters are accelerated only on prefix columns ^[12]. The replacement index is `(workspace_id, is_live, embedding)`, where `is_live` is a stored computed column so it cannot drift from the eviction timestamp it comes from ^[13]. The capability probe now reads the query plan instead of trusting that an index exists ^[14]. The author also found that vector indexing works on CockroachDB's free Basic tier: on v26.2.1, on 2026-08-03, the cluster setting read true and `CREATE VECTOR INDEX` completed ^[15]. Both findings went into a CockroachDB docs issue ^[16].

A coverage verdict cannot catch every failure. The mixed-embedder bug threw nothing ^[17], so the fix went in at write time: seeding now refuses an embedder that differs from recall's, and says why ^[18]. The French case is harder. Without cloud keys, recall uses a small local embedder that matches words, so a French question about an English memory truthfully reports a full search with nothing relevant ^[19]. "The receipt proves the search ran. It can't prove the search was clever," the author wrote ^[20]. Matching meaning is left to the hosted embedder, Titan on Bedrock ^[21].

The evidence for all of this is one developer's build. Throughline was one of 444 entries competing for three prizes ^[22]. The winner, Anchor, was also an on-call agent with an Unknown verdict, and it commits memory and actions in the same transaction ^[24]. The author kept Throughline's memory layer independent of the database on purpose, and offered a test count where the winners offered measured comparisons against a baseline ^[23].