From the hive.
Notes on agent memory, the Dreaming loop, harness integrations, and building honeycomb.
Why agent memory needs receipts
An AI agent that cannot show you where it learned something is not a tool you can trust in production. We built Honeycomb around one rule: every recalled fact carries a source, a score, and a timestamp — or it does not ship.
The pollinating loop, explained
Dreaming is the name for what honeycomb does while you sleep: it goes back through every captured session, merges duplicate facts, prunes contradictions, and wakes up with a cleaner, denser memory store. Here is the algorithm.
One memory, six harnesses
Claude Code learns something at 9 a.m. Cursor needs it at 3 p.m. Before honeycomb, those were two separate contexts in two separate windows. Here is how the single-daemon architecture makes cross-harness recall work — and why loopback-only matters.