Graph retrieval
Adds the missing hop when the answer lives one relationship away, without replacing hybrid search.
Memory research
Memory research, in the open.
Recall quality, consolidation safety, and read-surface security are research problems, not settled ones. Each track below started as a tracked issue, shipped behind a flag, and was measured against a benchmark before it changed a default. The claims carry receipts: an issue, a pull request, and an A/B result.
The tracks
Memory worth
Tracks whether recalled memories actually helped, then spends the recall budget on the ones that earn it.
Consolidation operators
Makes every split, merge, and update explain what changed and which source memories it came from.
Deferred verdicts
A third path for ambiguous memories, so extraction can revisit them with more context instead of guessing.
Surprise buffer
Flushes the buffer on semantic novelty, so the moments worth remembering do not wait for the clock.
Reasoning traces
Preserves how the agent investigated and found the fix, not just the answer, so similar problems do not restart from zero.
Read-surface threat model
Studies adaptive extraction attacks and the controls that keep recall useful without making bulk exfiltration trivial.
Where the tracks lead
These tracks feed an in-repo paper that documents the methods and situates them against related work in agent memory. It is a working draft, not a published result, and this page will link the artifact as it firms up. For what is being built next, see the roadmap.