neuron-db is the memory — the real Rust core compiled to a ~1.5 MB WebAssembly module, running locally. Every message is stored and recalled by it; watch the neurons fire in the memory pane. The optional local Llama (WebLLM/WebGPU) is the voice on top — it never sees your history, only what the memory recalls.

The memory does the retrieving so the tiny model only has to narrate: it routes each question to the right document, recalls by meaning and association, and for a “walk me through…” question hands back an ordered span so even a 1B model recounts a chain in sequence. Paste something bigger than the context window and neuron-db simply absorbs it — the model still answers from recall.

It has feelings: give it a mood (“be playful”) or an opinion (“you love rust”, repeat to deepen), and that affect colors how it talks. Multiple chats each get their own isolated neuron-db scope. Tune everything in ⚙ Settings. Everything runs in your browser — the memory is WASM, the model is WebLLM (cached locally after first download). It can also fetch the live web: say “check the web for …”, “google …”, or ask the weather — the WASM itself calls a public API directly (Wikipedia / open-meteo / DuckDuckGo, no server) and the results are saved into neuron-db so you can recall them.