跪拜 Guibai
← All articles
Artificial Intelligence

OmniMemEval Puts 14 Memory Systems Through the Same Gauntlet

By MemOS ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Memory benchmarks have been a mess of incompatible configurations, making it impossible to know whether a system actually helps an agent or just burns tokens. OmniMemEval gives teams a single, reproducible yardstick that measures both accuracy and the context cost required to achieve it—the two numbers that determine whether a memory layer is viable in production.

Summary

The OmniMemEval framework runs 14 memory systems—Mem0, Zep, Letta, Supermemory, Hindsight, and others—through the same models, prompts, and judge criteria to produce directly comparable scores. It evaluates on two axes: Agent Memory (SWE-Bench, BrowseComp, LiveCodeBench, OmniMath, GDPVal) and User Memory (LoCoMo, LongMemEval, PersonaMem-v2, HaluMem, BEAM), using both OpenClaw and Hermes agents.

MemOS, the framework's reference system, posts the highest overall scores on LoCoMo (88.83), LongMemEval (89.20), and PersonaMem-v2 (40.58) while using far fewer context tokens than competitors. On Agent Memory tasks, it lifts OpenClaw's GDPVal accuracy from 34.48% to 62.07% and Hermes' SWE-Bench accuracy from 37.18% to 52.56%.

The evaluation also tracks cost: average turns and output characters per task. MemOS achieves its scores with shorter outputs and fewer interaction rounds, meaning the gains come from better memory retrieval and context organization rather than from stuffing more tokens into the prompt.

Takeaways
OmniMemEval tests 14 memory systems on 10 benchmarks across two agent frameworks (OpenClaw and Hermes) under identical models, prompts, and judge criteria.
MemOS achieves the highest overall scores on LoCoMo (88.83), LongMemEval (89.20), and PersonaMem-v2 (40.58) while using significantly fewer context tokens than rivals.
On Agent Memory tasks, MemOS lifts OpenClaw GDPVal accuracy from 34.48% to 62.07% and Hermes SWE-Bench accuracy from 37.18% to 52.56%.
MemOS averages 5,400 context tokens on LoCoMo versus 32,532 for Cognee and 17,395 for Mem0; on LongMemEval it uses 4,151 tokens versus 117,106 for graphiti-zep.
Three technical changes drive the gains: context-enhanced extraction that pulls in surrounding chunks before memory extraction, a temporal parsing module that normalizes relative time expressions into standardized timespecs, and a span-based Evolve pipeline that processes execution DAG segments instead of full trajectories.
The full evaluation code and configuration are open-sourced so any team can reproduce the results.
Conclusions

Most memory system vendors publish scores using different underlying models, prompts, and retrieval settings, which makes their numbers non-comparable. OmniMemEval's insistence on a single judge model and identical answer-model configuration across all systems is the first serious attempt to fix this.

The context-token numbers expose a dirty secret: several systems achieve passable accuracy only by shoving enormous context windows into the prompt. MemOS's 4,151 tokens versus graphiti-zep's 117,106 on LongMemEval is not a small gap—it is the difference between a system that retrieves and one that dumps.

PersonaMem-v2's focus on sensitive preferences, safety boundaries, and forgetting requests tests something most memory evals ignore: whether the system respects what it should not remember. A high score here with low context tokens suggests the retrieval is precise enough to avoid leaking sensitive information into unrelated queries.

The Evolve pipeline's shift from full-trajectory processing to DAG-based span batching is a cost play that matters more than the accuracy numbers. Running reflection and compression on every step of a 50-step task is economically non-viable; processing only the high-information-density spans makes continuous agent learning affordable.

Concepts & terms
Agent Memory vs. User Memory
Two evaluation axes in OmniMemEval. Agent Memory measures whether a memory system helps an agent complete complex tasks better over time (experience reuse, strategy evolution). User Memory measures whether the system builds and maintains an accurate long-term model of the user's preferences, history, and relationships across sessions.
Context Token cost in memory evaluation
The number of tokens a memory system injects into the prompt to achieve its accuracy. A system that scores high by stuffing 100k+ tokens of retrieved context into every call will incur high latency and API costs in production, making raw accuracy alone a misleading metric.
Timespec (temporal specification)
A normalized representation of relative time expressions like 'next Wednesday' or 'after the last meeting' that MemOS generates during memory extraction. The reranker later uses these timespecs to sort and weight memories by temporal relevance, rather than treating time as raw text.
Evolve pipeline with DAG-based spans
MemOS's approach to learning from agent execution traces. Instead of processing an entire 50-step trajectory end-to-end, it records a lightweight trace, splits the execution DAG into spans at state-change and feedback boundaries, and runs compression and reflection only on those high-signal segments.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗