Skip to content

LongMemEval — Hebb Mind vs Mem0

Mem0's headline LongMemEval metric is end-to-end QA accuracy (the official retrieve → generate → LLM-judge protocol). Mem0 does not publish session-level retrieval recall, so the head-to-head is on QA.

SystemQA accuracyReader LLMReader prompt
Mem0~85–94%*gpt-4o / variousheavily engineered (public)
Hebb Mind v0.1.679.0%DeepSeek-V4-Proneutral official reader

* Mem0's reported figure varies by source and setup — their own research reports ~93–94%; third-party reproductions land near ~85%.

How to read this honestly

A same-number comparison is misleading here, for two reasons:

  1. Reader prompt does the heavy lifting. Mem0's answer-generation prompt is public and heavily engineered — 13 numbered rules, a chain-of-thought scratchpad, and an explicit personalization / anti-abstention section. Hebb's 79.0% uses the neutral official reader prompt with zero benchmark tuning. We measured this lever directly on our own stack: swapping our restrictive reader prompt for the neutral official one moved overall QA +12.4pp and the preference category +53pp. Reader-prompt engineering alone moves the number by >10pp — independent of memory quality.

  2. Retrieval is not the bottleneck for Hebb. Mem0 doesn't report retrieval recall; Hebb's is near-ceiling — 99.4% recall@10, ahead of Zep's 95.5%. The evidence is almost always retrieved; what caps Hebb's QA is the (deliberately untuned) generation prompt, not the memory layer.

So Mem0's edge on this benchmark is reader-prompt engineering, not memory quality. We publish the untuned 79.0% as an honest floor; a comparably-engineered reader on top of Hebb's stronger retrieval (99.4% recall@10) would be expected to close the gap.

Sources: Hebb eval/reports/longmemeval/v3/run-16 (QA) and run-14 (retrieval); Mem0 memory-benchmarks, research.

Released under the MIT License.