Recall

One search entry point fusing keyword and vector search - and the three different ways your knowledge actually reaches an agent.

In Seamless, recall is the single search entry point an agent calls to retrieve durable knowledge - memories and notes - from the store. It fuses SQLite FTS5 keyword matching with float32 cosine vector similarity using reciprocal rank fusion, all inside one SQLite file with no external vector database. With no embedding provider configured, recall runs keyword-only rather than failing; when a configured provider is merely unreachable, it degrades the same way instead of erroring.

recall is also the only search tool. There is no separate keyword search, no separate semantic search, no "advanced" variant. One entry point, because a surface with three search tools is a surface where agents pick the wrong one.

What it searches

recall spans two halves of the store, and the default scope (all) covers both:

The work record is there because knowledge does not only get written as a memory. A task body carries the acceptance criteria and the record of what was already tried; a dropped task is often the only account of why something was not done. A trial carries expected-vs-actual. A session's findings are the handoff its author wrote for whoever came next. All of it used to be reachable only by an agent already knowing which list tool to call - which is to say, only by an agent that already suspected the answer existed.

Narrow it with scope: memories, notes, tasks, trials, sessions.

Two things follow from the work record being keyword-only. It carries no embedding: there is no file behind a task row and so no content-hash reconcile loop, and vectorizing on write would put a provider round-trip inside every tasks_add. And because fusion rewards an item that both retrievers rank, a work-record hit can never outrank a memory that matched both ways on the same query - the ordering prefers curated knowledge without needing a rule that says so.

Work-record hits carry a status (a task's status, a trial's outcome), which is what separates a step you can still claim from a decision you can only learn from.

How it works

Two retrievers run over the store and their rankings are fused:

• FTS5 keyword search. Exact terms, names, error strings - the things vector search is worst at. If you know a memory is called chroma-boot-race, this is what finds it.
• Vector similarity. Embeddings stored as float32 blobs in SQLite, compared by brute-force cosine. This finds the memory that is about your problem when you don't know its name. Knowledge kinds only - see What it searches.

Their results are combined with reciprocal rank fusion: each retriever's ranking contributes, and something both rank highly wins. Neither retriever gets a veto, which matters because each is confidently wrong in a different way. (RRF for agent recall defines the method and the constant.)

The result set is then packed into budgets.recall_budget_tokens.

About that brute-force scan

Cosine similarity is computed by scanning every candidate vector. No ANN index, no approximate nearest neighbours, no vector database. For a personal knowledge store this is simply fast enough, and it buys exactness and one fewer moving part. It is a deliberate trade, not an oversight - and it is the kind of thing worth saying out loud rather than hiding behind the word "hybrid".

The utility nudge

Every memory carries a utility score: a time-decayed record of query-gated demand. An explicit read weighs 3.0, a recall hit 1.5, a prompt match 1.0 - and being pushed into a briefing counts for exactly nothing, so exposure can never masquerade as usefulness. Scores halve every 14 days.

Recall applies it as a bounded post-fusion multiplier: at most +10%, strictly below the +15% favorite boost, so an explicit star always outranks an implicit hot streak - and neither can resurrect something the retrievers did not return. The ambient <seam-recall> injection uses the same score to pick which qualified matches fill its three slots - after its overlap and score floors, never instead of them. The briefing blends it with recency once a project's demand history matures - for the memory index's order, and for ranking constraints ahead of their full/compact tier split (Sessions & briefings). One signal, three surfaces, all bounded.

The recall triad

This is the part worth internalizing: searching is only one of three ways your knowledge reaches an agent, and they are easy to confuse because all three end with the agent knowing something.

The first two are ambient: they happen whether or not the agent thinks to ask. That is the whole point of the hooks. An agent that never calls a Seamless tool still starts with your constraints and still gets relevant memories surfaced when its prompt touches them.

The third is what an agent does when it wants something specific. Before you go looking for something in your own knowledge base by hand, ask an agent to recall it - that is faster and more accurate than making you search.

Read what was already injected before searching again. The briefing is in context; re-recalling it just spends tokens to learn what the agent was told.

Degradation is deliberate

Recall keeps a distinction that is easy to get wrong: a remote failure degrades, a local misconfiguration surfaces.

• The embedding provider is unreachable, rate-limited, or rejects the key → degrade. Recall falls back to keyword-only results. You get worse ranking, not an error, because a partial answer beats no answer when the network is having a bad day.
• The configuration itself is wrong → surface it. This is not a transient condition that retrying fixes, and silently degrading forever would hide a problem only you can fix.

The failure mode this avoids: a store that has quietly been keyword-only for three weeks because nobody noticed the embedding key expired.

Writing memories that get recalled

Recall can only find what was written to be findable. Since the description is the only text an index shows, and it is heavily weighted in matching, it is the retrieval surface - see Memory & notes.