FlyBrain-Pollard-VL-CNSv1

A fruit-fly connectome used as the memory of a vision-language model. The artifact is the brain -- 117 MB of connectome, synapse signs, trained adapters and gate. It attaches to a frozen stock Qwen2-VL-2B-Instruct and is never merged into it.

This is the same brain design as FlyBrain-Pollard-CNSv1, trained against a different backbone. Not a variant, not a second architecture -- the same trainer, the same recipe, no vision-specific code anywhere. That is the point of publishing it: the method is backbone-agnostic, and this is the evidence.

Measured

A six-letter string is stated once, buried under filler, and asked about far beyond the attention window. The model never sees more than 128 tokens at a time, so it cannot see the fact and the question together. Words are generated fresh for every sample and never reused.

FLOOR -- no brain, fact outside the window	0.0%
RECALL -- with the brain	100.0%
CONTROL -- a word the document never contained	0.0%
CONTROL -- brain read a different document	0.0%
live state, at any document length	11.2 MB

64 samples through pollard-brainverify, which carries the verified measurement construction. Cold start reached 100% at step 44 and held it flat across eight consecutive evaluations.

Honest scope -- read this before assuming what it does

This brain recalls TEXT exactly, on a model that can see. It does not recall images.

Exact recall works by storing token ids as bipolar signs, and a sign only has to stay on the correct side of zero to survive interference. Qwen2-VL's vision encoder is continuous -- image patches have no ids -- so there is nothing discrete to store. The brain sits on the LANGUAGE side of the model.

Because a VL model injects visual embeddings into the token stream, the brain does see image content flow past and can address on it. What it cannot do is reproduce it. Byte-exact image recall needs a backbone whose image tokenizer is discrete (VQ-style), where image tokens are real ids and the existing bit path applies unchanged. That is a backbone choice, not a brain feature.

Use it

pip install 'pollard-weights[flybrain]'

from pollard_flybrain import FlyBrain, load_backbone
from transformers import AutoTokenizer

M = "Qwen/Qwen2-VL-2B-Instruct"
tok = AutoTokenizer.from_pretrained(M)
model = load_backbone(M, device="cuda")          # frozen; the vision tower is untouched

brain = FlyBrain.load("FlyBrain-Pollard-VL-CNSv1.pt").bind(model, tok)
brain.feed(open("long_document.txt").read())
print(brain.recall(" Question: what is the secret word? Answer: The secret word is"))

The prompt is part of the experiment. A token is filed under the words immediately before it, so a query has to reproduce that context. Ask "what is the secret word?" alone and a brain measuring 100% measures 46% -- the memory is intact, the question arrives at the wrong address. Verify any brain with pollard-brainverify, which has the correct construction built in.

Train one for YOUR model

A brain is fitted to one backbone: the address matrix has that model's hidden size (1536 here) and the token codes come from its output embedding. bind() refuses a mismatch rather than returning confident nonsense. For any other model, train your own -- it is quick:

pollard-flybrain --train 900 --model <your-hf-id> \
    --connectome graph.feather --signs signs.npy \
    --probes corpus.txt --brain MyModel-FlyBrain.pt
pollard-brainverify --brain MyModel-FlyBrain.pt --model <your-hf-id> --filler corpus.txt

pollard-brainlanes reports which runtimes can host a brain on your machine.

Also true

• It remembers; it does not reason. Language, reasoning and vision come from the backbone.
• Single-fact recall is solved; multi-fact selection is not -- several facts in one document is ~72-80% and is the open problem.
• PyTorch path. Verified lanes: transformers, MLX, GGUF/llama.cpp, EXL3. Not every runtime can host a brain, and pollard-brainlanes tells you which can on your hardware.

Connectome

MaleCNS v1.0 (FlyEM/Janelia), CC-BY. The full volume is 188,778 neurons and 26,028,386 synapses; this uses the associative-memory core -- mushroom body and central complex -- pruned to 8,552 neurons and 300,880 synapses.