Reproducible benchmark

MenteDB vs mem0

A measured, reproducible head-to-head on real LongMemEval conversations. Both systems get the same long histories, the same extraction model, and the same embedder, so any difference is the memory system, not the setup. The result: MenteDB ingests roughly 7x fewer tokens and ~6x cheaper than mem0 at comparable accuracy.

How we tested

• Real questions. Items from LongMemEval-S in the long categories (multi-session, temporal-reasoning, knowledge-update), full ~200-exchange histories, no trimming.
• Same models for both. Extraction with AWS Bedrock Claude Haiku 4.5; answering and grading with Claude Sonnet 4.5; embeddings with a local fastembed model. Identical for MenteDB and mem0.
• Exact cost.Every Bedrock call each system makes is intercepted and its tokens summed. MenteDB's write path (store plus supersede) makes no LLM call and is deterministic.
• Validated grader.The judge is checked against a control set: it fails stale, hedged, wrong, and confabulated answers, and the answerer says "I don't know" on empty context, so answers follow retrieval, not the model's own knowledge.

Honest caveats

• Accuracy is comparable, not better. 3/5 vs 2/5 is a one-question difference on five questions, i.e. within noise. We do not claim an accuracy advantage. The robust, structural result is cost and speed.
• Small sample. Five questions, one run. The per-turn cost gap is stable and structural; scale the sample up if you want tighter accuracy numbers.
• Why the gap. mem0 issues multiple LLM calls per write (extract, then add/update/delete decisions); MenteDB does one extract-plus-contradiction call per turn and reconciles with deterministic supersede edges.