Research

# Memory systems should be measured in public.

Our reports document the datasets, evaluation methods, architecture, and results behind supermemory’s performance.

Public benchmarks

## State of the art across agent memory.

97%

### LongMemEval-S

Recall@20 across 500 questions and six categories.

#1

### LoCoMo

Long-horizon conversational recall across single-hop, multi-hop, temporal, and adversarial questions.

#1

### ConvoMem

Personalization and preference learning across extended conversations.

Evaluation report · May 2026

## LongMemEval-S

[Read the full report](https://supermemory.ai/research/longmembench/)

supermemory reaches 97% overall Recall@20 with aggregation and leads the strongest baseline in every category.

__LongMemEval-S Recall@20 with aggregation__
| Category                       | supermemory | Zep   | Full context |
| ------------------------------ | ----------- | ----- | ------------ |
| SSUSingle-session — User       | 97%         | 92.9% | 81.4%        |
| SSASingle-session — Assistant  | 100%        | 80.4% | 94.6%        |
| SSPSingle-session — Preference | 95%         | 56.7% | 20%          |
| KUKnowledge Update             | 100%        | 83.3% | 78.2%        |
| TRTemporal Reasoning           | 95%         | 62.4% | 45.1%        |
| MSMulti-session                | 96%         | 57.9% | 44.3%        |
| ALLOverall                     | 97%         | 71.2% | 60.2%        |

DatasetLongMemEval-S · 500 questions · six categories

RetrievalRecall@20 with aggregation

Judgegpt-4o with the benchmark’s question-specific prompts

Systems research

## Memory as a filesystem.

SMFS · xAFS benchmark

### Less context, better answers.

SMFS cuts cumulative token usage by 3× on Claude and 1.75× on Codex while improving task accuracy across 110 questions.

[Read the SMFS research](https://smfs.ai/)