A notebook bolted to a frozen model
Hebb is a writable memory layer that sits on a model you already run. It learns after deployment. The model's weights never move.
How a write works
- Allocate. A cue is compared against every page key by cosine similarity. The best free page is claimed. No labels, no schema.
- Write. Gradient descent runs for roughly twenty passes with everything frozen except that one page's contents. The same mathematics as training, clamped to a few kilobytes.
- Read. The model cross-attends to the addressed page at every layer, for every token — not pasted into the prompt, wired into the forward pass.
Why it compounds
Teaching a model one thing usually costs you another. Fast weights reach 0.667 on a task and end the stream at 0.072: the state that held the new skill is the state that held the old one, and the new write overwrote it. That is why adaptation is normally a cycle rather than an accumulation — teach, measure the damage, retrain.
A page is not shared. Every memory that gets one keeps it, so corrections add up instead of trading against each other: immediate accuracy 1.000 and retained accuracy 1.000 across a hundred sequential tasks, with no gap between the two. Nothing learned on the third task is paid for by the ninety-seventh.
The mechanism is bookkeeping rather than a better update rule. The fraction of memories still holding a page of their own predicts retained accuracy to within 0.0039, with no free parameters. Retention is not something the method does gradually better as it trains; it follows from whether a memory was given somewhere to live.
So it compounds until it runs out of room. That is the honest shape of the claim, and it is the better failure to have: capacity is something you can measure, price and engineer, where decay is a property of the method that you cannot. The live ceiling is the addressing layer rather than the pages themselves — across 198 rules the best false-merge rate is 0.353, so a new memory can still land on a page that already belongs to something else. The numbers are on the benchmarks page.
What it is not
Not retrieval
RAG stores text and puts it in the prompt. A page holds learned values that need not correspond to any sentence, and never enters the context window.
Not fine-tuning
No weight in the base model changes. Nothing is merged, redeployed, or version-pinned to a training run.
Not a cache
A KV cache is discarded when generation ends. A page persists, and is still correct a hundred tasks later.
Where it fits
The case for it is write frequency. Adapters can hold per-tenant state, but updating that state means a retrain cycle. If a correction has to survive into the next request rather than the next release, a retrain cadence is not slow — it is the wrong shape.
Deployment
Self-hosted, in your own environment, on your own model. Requests never leave your perimeter, and the inference runs on hardware you already pay for. The API is four endpoints.
Status: research prototype, pre-customer. The addressing layer is the live constraint and is reported honestly on the benchmarks page. If you are evaluating this for production today, start there rather than here.