Skip to content

LongMemEval — Hebb Mind vs MemPalace

MemPalace and Hebb Mind both report session-level Recall@5 on the full 500-question LongMemEval-S, computed the same way (does the top-k contain a memory from an evidence session?). MemPalace's benchmark is retrieval-only, so this is a clean, apples-to-apples retrieval comparison — same metric, same question set, same MiniLM-384 embedding.

SystemR@5EmbeddingLLM rerank?Notes
MemPalace raw96.6%MiniLM-384NoVerbatim session docs
MemPalace hybrid v2 (temporal + 2-pass)98.4%MiniLM-384No
MemPalace hybrid v3 + Haiku rerank99.4%MiniLM-384YesTuned on the reported set
MemPalace hybrid v4 (held-out 450q)98.4%MiniLM-384NoHonest non-overfit number
Hebb Mind v0.1.699.0%MiniLM-384Yes (bge-reranker-base)Production hook mirror, full 500

Hebb's 99.0% R@5 (R@10 99.4%, R@1 93.4%) sits at the top of MemPalace's range — matching its best hybrid+rerank configuration on the same MiniLM-384 embedding, and above its honest held-out figure (98.4%).

Source: Hebb eval/reports/longmemeval/v3/run-14/longmemeval.md; MemPalace benchmark page.

On overfitting

MemPalace's hybrid v1–v3 numbers were tuned on the same 500 questions they report; their held-out v4 (98.4% on 450 unseen) is the honest non-overfit figure. Hebb Mind does not train or tune on LongMemEval — there is no train/test split (we don't fit a model), so 99.0% is a single full-set run with default production settings (the same hebb.json a user gets out of the box).

Beyond retrieval

MemPalace reports retrieval recall only. Hebb additionally runs the official end-to-end QA protocol — see the main LongMemEval page (79.0% with the neutral official reader) and vs Zep / vs Mem0 for the QA comparison.

Released under the MIT License.