Reproducibility: The benchmark harness is in github.com/yoheinakajima/act…. The repo pins the LongMemEval submodule, Python version, reader settings, judge model, embedding model, lockfile, dataset checksums, per-run manifests, and token counts

ActiveGraph is not the highest public LongMemEval number. Its narrower position is that it is a low-context, non-generative-ingestion substrate result: 85.6% QA at 2.4k context tokens, with paired statistics, pinned artifacts, and a replayable event-sourced runtime underneath.

The nominal ActiveGraph advantage over dense turn-RAG comes mostly from multi-session and temporal-reasoning questions. The category where it loses is knowledge-update.

ActiveGraph is within noise of the gold-session oracle (-1.2 points, n.s.) at a small context budget. Session-level retrieval uses roughly 10x more context and scores much lower.

Retrieval method: - grabbed top-k turns from conversation via vector search and grabbed before and after turns from those until context was full, leveraging activegraphs native event storage and graph retrieval

[technical blog post] “Evidence Compilation Before Semantic Memory: ActiveGraph on LongMemEval-S” —— 🔍85.6% QA accuracy and 86.2% turn answer-in-context at 2,462 mean context tokens, with deterministic non-generative ingestion activegraph.ai/blog/evidence…