First tagged release of neuron-db — an associative memory you can run anywhere, and the flat-cost long-term memory for an LLM. Pure-Rust core, zero dependencies, compiles to WebAssembly; durable storage, encryption, an HTTP server, and an MCP server are opt-in.
▶ Try it in your browser — no server, no key: https://gary23w.github.io/neuron-db/lab.html
What's new in 0.3.0
🧠 The in-browser lab is reasoned-routed — the model thinks, neuron-db grounds
A local WebLLM model runs entirely in your tab with neuron-db (as WASM) as its long-term memory. Each turn: perceive a knowledge-gap signal (how much of the question memory already holds, phrasing-independent) → the model reasons and picks one action (answer · web_search · deep_research · store · set_rule) → neuron-db acts and the answer is grounded in the result. So it searches when it doesn't know instead of confabulating, working memory + a persistent focus keep multi-turn tasks on subject, and commands can't be mis-stored as facts.
🌐 Live web access from WASM — no proxy, no worker
The WASM declares a host_http import the browser fulfills with fetch, reaching public CORS APIs directly (Wikipedia, DuckDuckGo, open-meteo). A recursive deep-research crawl follows Wikipedia's link graph into neuron-db. Nothing leaves your browser; the production build omits the import.
⚡ New WASM ops + a ~100× recall fix
New mem() ops: assess (the gap signal in the core), recallscored, vars/delvar, dump/load, scopes. Removed an eager O(N) scan in recall() — value/assess/chain drop from ~6 ms to ~60 µs on a 5k-fact scope.
🎨 Site & demos
Balanced two-column hero, fixed a page-scroll regression, all eight interactive demos made responsive phone→desktop, and the lab defaults to the smallest fast model (~0.9 GB, one-time download then cached).
Core
Tiers: Neuron · PlasticNeuron (adaptive) · NeuronRouter (sharded) · NeuronDB (durable SQLite) · SecureNeuronDB (AES-256-GCM) · HTTP serve · neuron-mcp (stdio MCP). Measured vs a markdown-dump LLM memory: 100% multi-hop accuracy at flat ~1.1k tokens/turn; selective recall ~µs out to 1,000,000 facts; ~130× denser than a 1536-d vector store.
Tests: 135 passing across all features.
Build
./build.sh # native (sqlite + secure + server)
cargo install --path rust/neuron-core --features mcp # the neuron-mcp stdio server