The owned, verifiable memory layer for AI agents — on Walrus. Capture anything, recall it by meaning, and prove it was never tampered with.
const neurus = await Neurus.open("my-agent")
// Remember anything — extracted into neurons on Walrus
await neurus.note("Sarah is allergic to shellfish. I owe her the deck Friday.")
await neurus.addFile("./contract.pdf")
// Recall by meaning — two-stage retrieval, grounded + cited
const answer = await neurus.ask("what do I owe Sarah, and when?")
// Plug it into any LLM as a retriever
const passages = await neurus.retrieve(query, { topK: 5, mmr: 0.5 })
// Prove the memory was never tampered with
await neurus.makeVerified()
const { ok } = await neurus.verifyIntegrity()owned, verifiable, on Walrus
notes and facts extracted into linked neurons automatically.
PDFs, docs, and folders chunked and indexed for recall.
two-stage retrieval with a cross-encoder reranker.
it surfaces contradictions instead of guessing.
sleep-time reflection synthesizes what matters.
verifiable, access-controlled knowledge sets on Walrus.
Owned, verifiable, persistent memory your agents can act on — fast retrieval, grounded reasoning, and proof it was never tampered with.
Why bi-encoders aren't enough, and how we pair a cross-encoder with Anthropic-style contextual chunking.
Every memory is a neuron; links are synapses. How the layers fit — from ingest to proactive reflection.
Bodies through MemWal, files on Walrus, a manifest you can publish and verify — and why the index is a cache.
An OR-Set op-log with Lamport clocks and capability signatures, so agents converge and bad edits are ignored.
Every neuron is stored on Walrus, encrypted with Seal under your own keys. No company holds your plaintext — you own the data, and an agent only ever sees what you grant it, revocably.
Content-addressed on Walrus and anchored by a Merkle root on Sui — agents acting on a verified set refuse to run on memory that has been altered, and anyone can prove what was stored, and when.