W4M · Research Notes

Research Note · New results · 7 / 12

It sleeps, it remembers, it grows up

Note 2 showed a model that learns while you use it. This is what happened when we gave that learning a full life cycle — days of use, corrections, and a nightly sleep — and measured every step.

W4M Research · July 2026 · ~7 min read

87 → 100%
taught itself to a perfect score in four rounds — zero human data, replicated
1 day
to master a tool it had never seen (33 → 100%, permanent, old skills untouched)
+64 pts
added to a completely stock small LLM through a 3 MB thought-token adapter
The short version We gave GoM's continual-learning core a life instead of a benchmark: days of real use, corrections from a user, and a nightly consolidation pass we call sleep. Left to generate its own practice, it took itself from 87% to a perfect 100% in four rounds with no human lessons. Handed a tool it had never seen, one day of being corrected made the tool permanent — with every old skill still intact. Its memories follow the same forgetting curve psychology measured in humans, and the same fix works: rehearsing on a spaced schedule kept 81% of memories alive versus 32% without. And when we let the brain speak through a completely standard LLM — no retraining, no surgery, just a 3 MB translation layer — the pair scored higher than either could alone. In business terms: a component that gets better with use, learns your tools after it ships, and upgrades any stock model it sits next to.

Note 2 made one claim: a small trainable brain attached to a frozen base can keep learning while you use it. This note is about what happens when that ability runs for days at a time — when the system doesn't just adapt to one switched rule, but lives with a user: answering, being corrected, sleeping on it, and waking up permanently better. Everything below is from logged runs, and every headline number was replicated on independent seeds before it went in this note.

It teaches itself — to a perfect score

The strongest version of learning-from-use is learning with no teacher at all. We let the brain generate its own practice problems, pick the ones it found hardest — it can measure its own uncertainty — and study those, checking its answers against the task's rules. No human wrote a single lesson.

Starting at 87% on a held-out exam, it reached 100% in four self-taught rounds, and stayed there. Run again on a fresh seed: the same climb to the same perfect score. The mechanism matters more than the number — the system chose what to practice, verified its own work, and consolidated the gains, entirely on its own. That loop is the engine everything else in this note runs on.

It chose what to practice, checked its own work, and consolidated the gains — no human in the loop.

One day to learn a tool it had never seen

Every model ships knowing a fixed set of things. The question a real deployment asks is different: can it learn your things — the internal tool, the house rule, the name that didn't exist at training time — after it ships?

We handed the finished, locked system a tool it had never encountered, and taught it the only way a user can: by correcting its mistakes in normal use. Day zero: 33% on requests for the new tool. After one day of corrections and one overnight consolidation: 100% — and it stayed at 100% in the days after. The part we consider the real result: every previously known tool also stayed at exactly 100% throughout. Learning the new thing cost none of the old things. Replicated on a second seed with the same outcome.

Never-seen toolAccuracyOld tools
Day 0 (first exposure)33%100%
Day 1 (corrected in use, one sleep)100%100%
Days 2–3 (no further teaching)100%100%

Sleep: how corrections become permanent

During the day, the brain holds what you taught it in a fast episodic memory — correct it once and it recalls that exact situation with 99–100% reliability, instantly. Overnight, a consolidation pass replays the day's corrections into permanent weights while rehearsing old skills so they don't fade. In the morning the short-term store is wiped clean; the lessons remain.

We measured how much a night can hold: retention stayed lossless up to 800 corrections in a single sleep — more than any realistic day of use produces. And we measured what happens without the rehearsal half: memories decay on the same curve psychology has documented in people for a century. The fix is also the same one that works for people — spaced review. Re-consolidating on a schedule kept 81% of day-one lessons alive ten days later; skipping it left 32%. The system now knows not just how to remember, but how to keep remembering.

Growing up: days of wild language

The hardest test we ran looks like a childhood. We put the brain on material far outside its training — free-form, messy, unfamiliar text where it starts out badly — and let it live there for five days: answering, taking corrections, sleeping each night.

Its ability on that material roughly doubled across the five days, while its original skills held at about 85% — a stability-versus-growth balance we tuned deliberately and then replicated. An earlier version of the bedtime routine grew just as fast but forgot three times more; the difference between the two is now a measured dial, and we know its setting. A system that can be sent somewhere new, get meaningfully better by living there, and come home still itself is the deployment story continual learning has always promised.

Sent somewhere new, it gets better by living there — and comes home still itself.

Speaking through a stock model — and making the pair smarter than either

All of this judgment needs a voice. Our answer keeps the language model completely standard — no retraining, no modified layers, nothing invasive. The brain hands the model a short sequence of thought tokens: not words, but compact packets of its judgment and its confidence, which a 3 MB translation layer renders in a form the model natively understands. The model then answers in its own fluent voice, already knowing what the brain concluded.

On a benchmark of reasoning and tool-choice questions, the stock small model alone scored 26.7%. Listening to the brain's thought tokens, the same untouched model scored 90.8% — a 64-point lift from a component you could attach to any model you already run. More bandwidth carries more judgment: the transmission improves smoothly with the number of thought tokens, and an adaptive version lets the brain spend few tokens on easy questions and more on hard ones, at no loss.

One line in the logs stands out. On tool decisions, the pair scored 98.3% — higher than the brain's own 92.4%. The language model isn't just repeating what the brain tells it; it reads the question with its own eyes, and when the two disagree, it sometimes catches the brain's mistake. Two small systems checking each other outperformed both. That is the shape of the swarm thesis from Note 5, measured at its smallest scale.

SetupAccuracyBase weights changed?
Stock small LLM, on its own26.7%
Same model + brain's thought tokens (3 MB adapter)90.8%No
Tool decisions, pair vs brain alone98.3% vs 92.4%No
For the technical reader Self-teaching: the core generates fresh practice each round, selects by its own uncertainty plus a diversity mix, labels via task checkers, and consolidates; held-out exam 87% → 100% in four rounds, stable through round eight, replicated across two seeds. Post-ship tool learning: reserved-capacity binding taught purely through in-use corrections plus one consolidation pass; 33% → 100% in one day, held through day three, prior-tool accuracy 100% throughout, two seeds. Episodic recall of once-seen corrections: 99–100% (consistent with Note 2). Consolidation capacity: retention lossless to 800 corrections per pass. Spaced re-consolidation on an expanding-interval schedule: 81% retention at day ten vs 32% unrehearsed, two seeds. Out-of-distribution "childhood": five days of use on unfamiliar free-form text, target ability roughly doubled (2–2.5× across seeds) with ~85% retention of original skills; the growth-versus-forgetting trade is controlled by a measured rehearsal setting. Interface: thought-token sequences through a ~3 MB adapter into an unmodified small instruct model; re-verified at 27.8% → 91.8% on a 600-item mixed reasoning/tool-choice set, near-ceiling from 8 tokens upward once fully trained, with an adaptive-budget variant that matches fixed budgets at every measured length (training scheme under NDA); on the tool subset the pair exceeds the brain's own accuracy (98.3% vs 92.4%). The core remains a small linear-time, non-attention design; construction details are available under NDA.
Honest note Two limits, reported as measured. First: on unfamiliar text where many distracting facts crowd the relevant ones, accuracy still drops sharply — we traced this to a specific, fixable training-data gap (the training scenes were simply never crowded), and the corrected training run was in progress as this note was written. Second: the childhood's doubling is relative — material that starts very hard ends at "learnable," not "solved"; lived experience closes the gap faster than any architectural fix we tested, but it has not closed it yet. Every number above is from logged runs, replicated on at least two seeds where stated, and reported as measured rather than extrapolated.

Put the pieces together and the shape of the product is visible: a component that ships small, learns your tools and your corrections after it ships, keeps its memories on a schedule, grows where you deploy it without forgetting where it came from — and makes any stock model it sits beside meaningfully smarter, without touching a single weight underneath.

Methodology: all figures are from real, logged runs. Self-teaching (87% → 100% in four rounds) uses self-generated practice with checker labels and no human-authored lessons; replicated across two seeds. The never-seen-tool result (33% → 100% in one day, old tools 100% throughout) was taught exclusively through in-use corrections and consolidation, two seeds. Spaced-review retention (81% vs 32% at day ten) and the five-day out-of-distribution childhood (2–2.5× growth, ~85% skill retention) were each replicated on independent seeds. The thought-token interface result (26.7% → 90.8%; tool subset 98.3% vs the brain's own 92.4%) uses an unmodified instruct model with a ~3 MB adapter; the interface numbers were subsequently re-verified on a 600-item evaluation with a fully trained adapter: 91.8% vs 27.8% raw (+64 points), stable from 8 tokens upward. Numbers are reported as measured and are not extrapolated.

A model that improves with use — now with a life cycle to prove it

Self-taught to perfect, new tools in a day, memories on a schedule, and a stock LLM lifted 64 points without retraining. Qualified partners can request the full evidence.

w4m.ai