research

EvMind / EvAGI

Research into modular capacity, expert isolation and continual learning.

What has actually been run

Two experiments are published below. Both report zero measured forgetting, and both reach it the same way: each task is allocated its own disjoint parameters, so there is no shared weight left for one task to overwrite. That is the mechanism. It is why the result holds, and it is also the limit of the claim — isolation removes interference by construction rather than improving how a model retains what it has learned.

17K-parameter system, twenty-shape task

A single small model learning five shape classes in sequence. Chosen because a system that size can be inspected directly. About 31% of the expert pool is occupied by the end.

33M-parameter language model, five sequential skills

A small language model takes on sentiment, spelling, capitalisation, pluralisation and antonyms, then learns facts live. The live-learning check is repeated in a fresh session with an empty context and prompts audited for leakage. Roughly 5.5% of the pool is used for skills.

Both are single runs. No seeds, repeats or error bars are shown, neither task set is a standard continual-learning benchmark, and neither has been released for independent reproduction. The figures below are what was measured, not a solved problem.

The question

EvMind is the research programme on continual learning: how a model can absorb new domains without losing what it already knows. EvAGI is the experimental system associated with that work, where live-learning behaviour is demonstrated rather than only measured.

Studies modular capacity and routing between specialised areas
Examines expert isolation as a way to limit interference between learned domains
Keeps experiment logs, metrics and plots so results stay inspectable

What has been run

Experimental research with internal reports, experiment logs and metric plots. Two published runs are shown below: a single 17K-parameter system on a twenty-shape task, and a 33M-parameter language model on a five-task continual-learning sequence with a cross-session live-learning check. Quantitative claims are only published when the run, baseline and evaluation protocol can be stated alongside them. No public code or artefacts are released for this branch.

Evidence

Captures from the current build and runs. Slots marked as pending are not published as proof — they are waiting on a sanitised capture.

Eight-panel figure from a TinyStories-33M continual-learning run: final accuracy and average forgetting per task for a full fine-tune baseline against EvAGI, the occupied share of a 12,288-neuron pool, training and inference compute, accuracy per thousand trained parameters, a capacity curve, per-task final accuracy, and a cross-session live-learning check. EvAGI reaches 100% final accuracy on all five tasks with 0% forgetting, occupying 5.5% of the pool for skills plus 6.6% including two live-learned facts. — EvAGI on TinyStories-33M (68.5M total parameters): accuracy, forgetting, pool occupancy, compute and a cross-session live-learning check, from a single run.Read this as one experimental configuration, not a solved problem. Forgetting is zero because each task is allocated disjoint parameters — isolation removes interference by construction. The task set is five narrow skills on a small synthetic corpus, not a real-world benchmark, and this is a single run with no variance shown.

Eight-panel figure from a 17K-parameter continual-learning run over a twenty-shape classification task: final accuracy, average forgetting, expert pool occupancy, training and inference compute, accuracy per thousand parameters, a capacity curve against active weights, and per-task final accuracy across five shape classes. The EvAGI variants reach roughly 98% accuracy with 0% forgetting while occupying about 31% of the pool. — EvAGI on a single 17K-parameter system: the same measurement set on a twenty-shape classification task.This is a synthetic twenty-shape task at a scale small enough to inspect directly. The zero-forgetting result again follows from disjoint expert allocation, and a pool that reaches only about 31% occupancy would need to grow substantially before this scales.

Schematic of expert isolation: an input token stream enters a router, which dispatches to a row of visually disjoint expert blocks, each paired with its own separate parameter pool and separated by dashed boundaries so that no parameters are shared. Expert outputs then merge into a single result stream. — Expert isolation and routing. Schematic of the intended structure — not a capture of the running system, and not evidence about any measured result.

Source and documentation

Public artefacts for this branch, checked before publication.

DocumentationPublic docs — no source code

Where this connects

Branches share infrastructure and people; the relationship is association, not a dependency chain.

EvAgent
Available

EvMMFPS
Experimental

EvMem
In development

EvStudy
In development

EvSafety
Experimental

EvAgency
Active