Context Swarm Memory beats Hindsight on BEAM 100K, uses 38% fewer tokens

Context Swarm Memory (CSM) is a new open-source agent-memory system that uses bounded read-only memory shards, query routing, probe/recall/synthesis, cited packets, and explicit Committer-gated writes. In a head-to-head comparison against the accepted Hindsight artifact on the BEAM 100K benchmark, CSM achieved an AMB score of 0.757573 (342 out of 400 correct), surpassing Hindsight's 0.733658 (326 out of 400). Additionally, CSM uses 38.2% fewer answer-visible context tokens, which could reduce inference cost and latency in downstream agent calls.

However, the trade-off is speed: CSM's average retrieval time is 29.23 seconds versus 6.23 seconds for Hindsight. The developer acknowledges this is not an official leaderboard result—only a local accepted-artifact comparison at the 100K scale. They are seeking community feedback on how to strengthen the evaluation methodology before attempting independent replication or official submission to BEAM. The repository and reproducibility notes are publicly available on GitHub and a dedicated site.