Early in Bourdon’s development we wired one of our first agent integrations, ingested a real project’s worth of memory, and asked it a recognition-style question: what do you know about this project?
The answer was correct. The ingestion had worked, the entities were clean, the context was all there. And it still felt wrong — the agent searched first, then answered from gathered notes, like a librarian reading you their findings. The data layer had succeeded. The behavior layer had not.
Two different kinds of success
That day we started treating those as separate, measurable things, because they fail separately:
- Data-layer success: the memory is ingested, structured, and retrievable. Most memory products stop here, and most benchmarks only measure here.
- Behavior-layer success: the first response reads like recognition — immediate, natural, already oriented — while deeper context hydrates in the background.
A system can pass the first completely and fail the second completely. Ours did. Writing that down as a product finding, rather than shrugging it off as preference, is probably the single most consequential decision in the project: it made runtime timing a first-class workstream instead of a nice-to-have on top of ingestion quality.
What changed because of it
The concrete work that came out of that finding: an L0 alias map so known names resolve immediately to known identities without a visible lookup; automatic L1 hydration so a recognized entity surfaces its essentials — path, status, milestones — while the reply is already underway; and evaluations that score response timing style alongside context quality, so a regression into search-first behavior fails a check instead of slipping through.
We also committed to recording recognition failures as journal artifacts. Misses like this are the most useful evidence the thesis produces — more useful than the successes, honestly, because each one marks the exact edge where structured memory stops feeling like being known.
The honest caveats
This was a bounded finding on our own integrations, not a benchmark of anyone else’s. Different agents host different constraints, and some surfaces give an integration no control over first-token timing at all. Where that is true, the recognition-first behavior degrades to very fast retrieval — better, but not the full promise, and we say so per adapter rather than rounding up.
The raw entry behind this note, with its timeline and decisions, is in the findings journal. The architectural argument is in the previous field note.