Lab notebook · agent infrastructure
Working notes and tools for agent infrastructure: memory that can be audited, routing that runs without a hosted provider, and systems built to say something when they stop working.
Everything here is software that exists and runs. Where a claim has a number behind it, the number was measured.
Retrieval and memory for agents: vector and hybrid search over LanceDB, reranking, hierarchical summarisation, and a decision-context graph that records why a memory exists and what it rests on.
npm install @yamo/memory-mesh source →A typed memory graph over beads memories, as a pure layer — beads is never modified and every link lives as plain text inside the memories themselves, so it survives the tool being deleted. Admission control before writing, health checks, and an observation ledger supplying the timestamps the store does not keep.
git clone https://github.com/scgoetsch/bd-memgraph source →A daemon that routes work across local models by intent and difficulty, persists every dispatch with its route and outcome, and consolidates memory on a loop. Routing only learns from outcomes it can prove were real.
An earlier attempt at the same problem: hash an agent’s reasoning block, pin the body to IPFS, anchor the hash on Ethereum. The contracts deployed and the explorer worked, but the anchors ended up being the system talking to itself, and the immutability the pitch rested on was never really tested by a second party. The question it was asking — can you verify why an agent decided something — turned out to be answerable locally, with typed memory lineage and dispatch records, and without asking anyone to trust a chain. Retired rather than maintained; the repositories stay up as a record, and the full account is in the field report.