Blog
Building the memory layer
Notes from inside Neurus — retrieval, architecture, storage, and multi-agent memory on Walrus.
Beyond bi-encoders: cross-encoder reranking and contextual retrieval
MemWal gives us fast, broad recall — but bi-encoders are weak at precision ordering. Here is how we pair a cross-encoder reranker with Anthropic-style contextual chunking to find the right neuron, first.
Read →The architecture of Neurus: an owned, verifiable memory layer on Walrus
Every memory is a neuron; links between them are synapses. Here is how the layers fit together — from ingest to retrieval to proactive reflection — and why the whole thing lives on Walrus.
Read →Wiring it together: Walrus, MemWal, and your owned memory
Neuron bodies go through MemWal; files go straight to Walrus; the map is a manifest you can publish and verify. Here is how the three layers connect — and why the index is a cache, not the source of truth.
Read →Conflict-free shared memory: implementing CRDTs for multi-agent sets
When two agents write the same knowledge set at once, last-write-wins clobbers. Here is how we built an OR-Set op-log with Lamport clocks and capability signatures so they converge — and unauthorized edits are simply ignored.
Read →