The server injects up to five project notes before the agent takes its first turn

`_memory_preseed(query, top_k=5)` calls `memory_search` on the server, then formats each hit as a line carrying the note name, its description and a snippet, with a star prefixed when the note has `priority` set ^[10]. The model sees list order and the star; the code shown does not pass relevance scores through. If the hit list comes back empty the function returns an empty string ^[10].

The design came out of a compliance failure. A session reimplemented an error format the team had standardised weeks earlier, because nothing told it the decision existed, and the decision was written down in a markdown note the session never opened ^[3]. The first fix told each session to "Start by calling the memory_search tool to load any relevant project context" ^[4]. "And it mostly works. Mostly," the author wrote on dev.to ^[5]. Stronger wording, moving the instruction last, and tool descriptions that begged came next ^[7], and then the conclusion: "I was trying to make a non-deterministic thing reliable by asking harder. Wrong layer." ^[8] The post does not report how often a session actually skipped the call, so the one-in-nine framing is an argument about odds rather than a measurement, and the author's stated concern is that a skipped read is silent ^[6][21].

The release ships an end-to-end test that seeds a throwaway instance with two notes whose facts appear nowhere in the code, against a workspace README deliberately wrong about all three ^[15]. The spawned session is asked "what port does staging run on, which host do we deploy to, and with what command?", and passing means answering 6443, callisto and `make ship-v2` ^[16]. The transcript records the session saying it answered "without needing to search the codebase" ^[17]. A session under the old instruction that did call `memory_search` would answer the same three values. So the test verifies that server-side recall reaches the model, and comparing the two designs would mean spawning the same task both ways and counting the misses.

`top_k` defaults to 5 and the cockpit keeps about nine sessions alive ^[10][1], so a full round of spawns puts up to 45 note lines into first messages ^[20]. Each line is a note name, a description and a snippet ^[10]. Whether that is cheap in your setup depends on how long your descriptions run and whether the injected block sits inside a prefix your provider caches.

The store is the part I would copy. Notes are markdown with YAML frontmatter, the same files Claude Code already writes, one fact per file, with the `description:` doubling as the embedding text and an optional `priority: high` ^[18]. The index is incremental sqlite with the vectors in the same store, no pgvector and no ANN library, and the author reports that dot-product over a warm in-process model is instant at a few hundred notes ^[19]. For that to transfer you need the same conditions: a few hundred notes, one process, and the embedding model resident in it.

The convention is one fact per file ^[18]. The example injected block shows `[staging-stack]` carrying the staging port and the deploy host on a single line ^[11]. Because the description is the embedding text, a note holding two facts is only retrieved by whichever one the description mentions.