Hey HN.

Open-source version at: https://github.com/dat999zx/knowl

The number is FactConsolidation single-hop at 262K context on MemoryAgentBench. On multi-hop Knowl scored 0.07 against an all-time ceiling of 0.14.

Write-up of the approaches that failed first: https://blog.knowl.cloud/the-story-of-knowl

how do concurrent writes work if two agent sessions are running in teh same project?

cool stuff