Context Heavy: What the LoCoMo Result Shows

Context Heavy reached 71.2% recall@10 and 60.1% nDCG@10 on a retrieval-only, 304-question LoCoMo slice. In the same run, supermemory reached 63.3% recall and 45.4% nDCG, gbrain 58.0% and 38.2%, and mem0 20.6% and 17.4%.

What was measured

The experiment used two LoCoMo conversations, isolated by conversation, with neutral titles and k=10. It evaluated retrieval ranking: not final answer quality. Context Heavy and gbrain used the same BGE-M3 embedder; hosted systems ran as their shipped products allowed.

Why the system ranked well

Context Heavy combines lexical retrieval, vector similarity, reciprocal-rank fusion, graph context, and a cross-encoder reranker. That design is intended for questions where a paraphrase and a relationship both matter. The benchmark also includes temporal update and abstention cases in its custom suite.

The trade-off

The Context Heavy p50 was 1,708 milliseconds, compared with 713 milliseconds for supermemory. The remote reranker accounted for much of that cost. A ranking lead with a latency penalty is a design trade, not a universal win.

The caveat

Full LoCoMo is materially larger than this two-conversation slice. These retrieval scores are also not comparable with published answer-judge scores. The correct conclusion is narrow: on this reproducible slice and metric set, Context Heavy led the compared systems on ranking quality. CH-Bench publishes the adapters, caveats, and commands required to inspect the result.