W4M · Research Notes

Research Note · New results · 8 / 12

The swarm, measured

Note 5 was a thesis: many small models, learning together, growing as one network. This note is the first data — small in scale, exact in what it shows.

W4M Research · July 2026 · ~5 min read

45.6 → 100%
every member perfect after one overnight memory sync — on material only teammates had seen
×3
the result replicated three times, three brains each run
0 → 100%
one brain's experience transplanted into another — the recipient masters what only the donor lived
The short version We split a workload across three copies of the continual-learning brain from Note 7, so each lived through a different third of the problems. Alone, a brain that had seen only its own third scored 45.6% on the full exam. Then we did what frozen models cannot: synced their experience overnight, the same consolidation each brain uses for its own memories. By morning, every member scored 100% on everything — including material only its teammates had ever seen. Replicated three times. Experience also moves as a file: transplant one brain's memories into a fresh copy and it masters the donor's material immediately, 0 to 100. The honest caveat is real and measured too: shared learning is only as good as what each member perceives — feed the network noisy inputs and the gains shrink until perception is fixed. In business terms: a fleet where every unit's experience becomes every unit's skill, overnight, without a datacenter in the loop.

Note 5 sketched the destination: millions of small models on the edge, each learning from its own use, sharing what it learns so the whole network compounds. A thesis needs numbers. These are the first — deliberately small in scale, three to five brains rather than millions, because at small scale you can measure exactly what sharing does and control everything else.

Three lives, one mind by morning

The setup is simple to state. Take three copies of the same continual-learning brain. Give each a different third of a task world to live in — different problems, different corrections, no overlap. None of them ever sees the full picture. Then, at the end of the "day," sync their experience stores and let each brain run its normal overnight consolidation — the same sleep mechanism from Note 7, just fed with the whole team's day instead of its own.

Before the sync, a brain tested on the full exam scores 45.6% — it knows its third and little else. After one sync and one sleep, all three members score 100% on the complete exam. Not the union of the team held somewhere central — each individual brain, carrying the whole team's competence in its own weights. We ran the experiment three times; it landed the same way each time.

Each unit's experience became every unit's skill — overnight, in the units themselves.
SetupFull-exam accuracy
One brain, alone, having lived one-third of the world45.6%
Same brain after one overnight team sync100%
Its two teammates, same sync100% each

Experience is a file you can hand to a new unit

Because a brain's memories live inside its checkpoint — no external database, nothing server-side — experience is portable by construction. We took the memory store of a brain that had lived through a set of problems and transplanted it into a fresh copy that had never seen them. The recipient went from 0% to 100% on the donor's material, immediately, before any retraining of its own.

Practically, that is onboarding for machines: a new unit joining a deployed fleet does not start from zero — it starts from everything the fleet has already lived. And the speed benefit is measurable even day to day: in a five-member trial, the swarm reached full mastery of a shared world a full day before a single brain working the same world alone.

The honest limit — and why it points at the fix

One experiment in this series failed on purpose, and it is the most instructive. When we ran the sharing loop on inputs that each member perceived noisily — garbled versions of what users actually said — the collective gain collapsed to a few points. Shared learning amplifies whatever the members actually experienced: clean experience compounds, noisy experience doesn't. The network is exactly as good as its senses.

That is not a surprise ending; it is a design requirement we now have a number for. It is also why the perception work in Note 7 — the brain reading its inputs more faithfully, and the crowded-scene training fix noted there — sits upstream of the swarm on our roadmap. Fix what one unit perceives, and the sharing math from this note multiplies it across the fleet for free.

For the technical reader Fleet sharing: three same-base continual-learning cores, disjoint thirds of a task world, corrections collected per member during the day; experience stores unioned nightly and consolidated by each member's standard sleep pass. Full-exam accuracy: 45.6% solo-slice baseline → 100% for every member post-sync; replicated in three independent runs. Transplant: donor memory store loaded into a fresh same-base copy → 0% → 100% on donor-lived material with no recipient retraining. Five-member trial: swarm reached exam saturation one day ahead of a single-brain control; in small rule-worlds saturation then caps further compounding — larger worlds are the open scaling test. Perception dependence: the identical sharing loop run through a noisy input pipeline yielded roughly a three-point collective gain — shared learning is bounded by member-level input fidelity, which is the coupling to the perception work in Note 7. The memory store is part of the model artifact itself (no external database); sync is a file-level operation. Core and memory-organ construction details are available under NDA.
Honest note These are small-fleet results — three to five members on controlled task worlds, built to isolate the sharing mechanism, not to demonstrate internet scale. Two limits are measured and reported: in small worlds the swarm's advantage is speed-to-mastery rather than unbounded compounding (the world saturates), and the gains depend directly on how cleanly each member perceives its inputs. The millions-of-units picture from Note 5 remains a thesis; what this note establishes is the mechanism it rests on — experience that moves between units and becomes skill overnight — with replication.

The swarm story is often told as a metaphor. This note is the part that isn't: individual experience, made portable and shared, measurably becoming the whole team's competence — with the units themselves, not a datacenter, carrying every bit of it.

Methodology: all figures are from real, logged runs. The fleet result (45.6% → 100% for all three members after one nightly sync) was replicated across three independent runs with disjoint task slices per member. The transplant result (0% → 100% on donor-lived material) moves the donor's memory store into an otherwise fresh same-base model. The five-member speed result and the noisy-pipeline collective result (~three-point gain) are single controlled runs reported as measured. No result in this note is extrapolated beyond the fleet sizes tested.

Note 5 was the thesis. This is the first data.

Experience that moves between units and becomes skill overnight — replicated. Qualified partners can request the full evidence.

w4m.ai