Graft gates coding-agent recall behind a STRONG, WEAK or MISS verdict

Verified recall gates the answer. It hands back one candidate only when multiple signals support the match, and the field the caller branches on is a label: STRONG, WEAK, or MISS ^[5]. Each label carries a prescribed action. A STRONG hit is evidence, and the agent is expected to read the returned context and decide whether it fits ^[6]. WEAK should send the agent to investigate, and MISS tells it to solve the problem normally ^[7]. A caller writes three paths where a top-k endpoint lets it write one ^[22].

Underneath, two rankings run and then get merged. Dense embeddings catch paraphrase, while BM25 keeps the exact anchors that carry disproportionate meaning in engineering work, among them a configuration key, an exception type, a table name and a package version ^[8]. The post proposes reciprocal-rank fusion for the merge, on the grounds that it blends independent rankings without pretending either is universally correct ^[9]. Ask for help with a Docker healthcheck and similarity alone may return memories about container startup or service discovery, which the post treats as leads ^[3].

The store has to be the right shape for any of this to work. A codebase index is built to answer which documents might match a query; agent memory has to answer what the agent learned that would be expensive to rediscover ^[11]. The entries the post asks for are compact and explicit: a root cause with its fix, an architectural choice with the reason behind it, a constraint that rules out an appealing approach ^[12].

When a decision changes, the old entry stays in the store. Deleting it loses context and leaving it untreated risks a stale answer, so the post marks the old learning superseded and links it to the newer one. History stays inspectable, and retrieval favours what is current ^[13].

Then there is the threshold. The costs of being wrong are asymmetric. For an agent that proposes code changes, "a plausible but stale memory can send work down the wrong path", the author wrote ^[15]. For an agent doing background research, a false negative is cheaper, because it can simply search more broadly ^[16]. The post's rule follows from that: confidence should be exposed as part of the contract ^[17].

The evaluation loop is built out of failures: paraphrases of the same historical problem, near misses on a different version or environment, contradictions and superseded decisions, exact-identifier queries that vectors may dilute, and small noisy stores where every result appears relevant ^[18]. "It is to learn when the system should stay silent," the author wrote of the goal ^[19]. The post lists those categories without reporting how Graft scores on any of them ^[23].

For the gate to be worth the wiring in someone else's stack, two things have to hold. The store needs entries with enough structure that a second signal has something to check ^[12], and the calling agent has to do something different on WEAK than it does on STRONG ^[6][7].