Plastic Hippocampal Layer: memory in synapses

Each Tell is its own 512-token experience, padded with web text. The frozen SmolLM2-135M language model can only attend within it. While it reads, a local Hebbian rule writes into fast synaptic matrices, with no embedding store. Each Ask is a fresh context that can't see anything you told it. If the answer comes back, it came from the changed synapses. The no memory column is the same model with the fast synapses zeroed.

Experience (written to synapses)

Synapses are empty; nothing has been experienced yet.

Question (fresh context)

With memory

No memory (synapses zeroed)

What it was trained on (use these formats; any names work)

This is a research prototype: a frozen SmolLM2-135M plus a meta-trained plastic hippocampal layer (DG→CA3 sparse codes with a three-factor Hebbian write). It runs on RunPod Serverless, so the first request after the GPU has been idle can take about 20–60 s.