← Blog
ArchitectureJun 5, 2026·7 min read

The architecture of Neurus: an owned, verifiable memory layer on Walrus

Neurus is the intelligence layer over the Walrus data economy. Its spine is a single abstraction: the neuron — one stored fact, file chunk, person, commitment, or synthesized insight, indexed in MemWal and stored on Walrus. Neurons are joined by typed synapses (about, derived_from, promised_to, reflects_on), so your memory is literally a graph, not a flat table.

One engine, clean layers

Everything funnels through a single Memory engine, with thin layers around it:

ingest/    note · file · dir · walrus   → turn anything into neurons
core/      neuron · memory · sets       → the engine + knowledge sets
retrieval/ rerank · bm25 · rrf · mmr    → two-stage precision recall
reason/    answer · brief               → grounded, cited, conflict-aware
proactive/ reflect · surface           → sleep-time insight + interruption calculus
storage/   walrus · memwal             → owned, encrypted persistence
access/    seal        integrity/ merkle → revocable + tamper-evident

The unit a user works with is a knowledge set: a named bundle of neurons with its own namespace, manifest, visibility, and integrity tier. Your personal memory is one set; a company’s documentation indexed from Walrus is another. The same engine serves both — only the trust tag differs.

Reactive and proactive, in one graph

The reactive path is RAG done carefully: recall → cross-encoder rerank → an answer that uses only retrieved neurons, cites them, and surfaces contradictions instead of guessing. The proactive path is the differentiator — a sleep-time reflection process reads recent neurons, synthesizes higher-level insight-neurons(“three people are blocked on the Q3 deck”), scores their importance, and an interruption calculus decides what is worth your attention now. The clock is a trivial trigger; the intelligence is in the synthesis and the restraint.

Three surfaces, one core

The same engine is exposed as a TypeScript SDK (Neurus.open(set).ask(...)), a CORS-enabled HTTP API (/v1/ask, /v1/retrieve, …) so any language plugs in, a CLI, and a memory inspector that shows the neuron graph and the exact recall spans behind every answer. Retrieval quality, verifiability, and ownership are all properties of the core — not bolt-ons.