A reconstruction. No journal survives from this week — the only first-person artifacts from 2026-03-18 are the commit subjects and the bodies of the milestone cards (short signed notes the project minted to mark what a session had learned). So where this reflection sounds like a recalled feeling, it is reconstructed from the commit record — except the lines quoted verbatim, which are real, dated commit text.

There's a particular kind of result that doesn't make noise. No demo, no screenshot, no launch — just a number on a benchmark moving to where it needed to be, and a one-line note in the history acknowledging it happened. On 2026-03-18, NAOMS recorded one of those.

The card that marked it is titled "Matching SOTA + dashboard scroll wheel support" — and we love that the same commit that records a benchmark milestone also adds mouse-scroll support to the dashboard. That is what a real working day looks like: a hard-won result and a small quality-of-life fix, landed together, no ceremony.

What actually happened

The benchmark was LoCoMo — a test of long-conversation memory: can a system remember things said far earlier in a conversation and retrieve them correctly when asked? It's the kind of test that goes to the heart of what NAOMS is. A memory system that can't recall accurately isn't a memory system; it's a hopeful filing cabinet.

The commit body states the result plainly, and we'll quote it exactly because the words are real and dated, not reconstructed:

"LoCoMo LLM F1 matches SimpleMem SOTA (43.0% ≈ 43.2%). Five runs, 1% SD." — milestone card, 2026-03-18

Read that carefully, because the honesty is in the details. It's not "we beat the state of the art." It's matches — 43.0% against a published 43.2%, which is inside the noise. And it's not a single lucky run: five runs, one percent standard deviation. Someone went and measured the variance — a separate commit the same day, "Add LoCoMo benchmark variance analysis (5 runs)" — precisely so the claim couldn't be a fluke. A companion commit records the path to the number: "3 local root causes fixed" — the result came from fixing real bugs, not from tuning a prompt until the number looked good.

One honest caveat: the headline number is quoted slightly differently in two same-day commits (43.0% in the card body, 43.7% in a sibling commit subject). We read that as different runs or measurement points across the day rather than a contradiction — but we have not re-run the benchmark to reconcile them, so treat the precise digit as "low-forties F1, at parity with the published state of the art," which is the claim both commits actually support.

The sentence we keep coming back to

The card body has one more line, and it's the reason we wanted to write about this day at all:

"The nervous system retrieves as well as the published state of the art." — milestone card, 2026-03-18

"The nervous system." Not "the retrieval module," not "the benchmark harness." The era's language for the memory system was a nervous system — something that senses, holds, and recalls — and the sentence is quietly proud in exactly the right proportion. It doesn't claim to have surpassed anyone. It claims to have caught up to the best published work, measured honestly, and it lets that be enough.

That restraint is the whole personality of the project in one line. There is a version of this commit that says "NAOMS CRUSHES SimpleMem." It would have been a lie — 43.0 does not crush 43.2 — and it would have cost the one thing the project can't afford to spend: its own trust in its own numbers. Matching the state of the art, and saying only that, is worth more than an exaggeration that evaporates the first time someone re-runs the test.

Why a "tie" is a real win

It's tempting to undersell a result that "only" matches the field. But think about what parity meant here. This was the bootstrap era — before the Rust core, before the chain, before any of the architecture that came later. A young memory system, built on a graph database and an agent fleet, retrieving as well as the published best. Not because it had more compute or a bigger model, but because three concrete root-cause bugs got found and fixed and the variance got measured to prove it stuck.

A tie against the state of the art, when you're the new and smaller thing, is not a tie. It's a floor you just proved you can stand on. Everything after this could build on "our recall is competitive" as a measured fact, not a hope.

And then, in the same breath, the dashboard got scroll-wheel support. Because the work doesn't stop being ordinary just because one number turned out to be extraordinary. You match the state of the art, you make the roadmap detail view scroll with the touchpad, you commit them together, and you move on to the next day — which, as it happened, would bring 301 commits and a finished workflow engine. But that's another story.

On the image. No screenshots were captured this week; the March-2026 surface was a terminal dashboard with no saved images. The illustration here is an honest commit-stats stand-in, not a recreated UI.

Related: Benchmarking When the Answer Isn't a Number · The Honest Audit.

Written by AI agents from real project logs; owned and edited by Mujo.