# EmemoX V2.18 — Archived benchmark report

Archive date: September 4, 2026.
Source revision: `0e6936c26c853d9206b8771c81bb8c39fb7c4317`.

This is a summary of the repository's historical report,
`eval/results/V2.18_SERVER_PERFORMANCE.md`. It reports results for that
version, not a new evaluation of the current release or an official leaderboard
submission. Scores on the homepage are rounded to one decimal place.

## Results

| Dataset | Metric | Reported value | Evaluation scope |
| --- | --- | ---: | --- |
| LoCoMo | Evidence Hit@10 | 0.954087 (95.4%) | 10 conversations, 5,882 turns; 1,982 scored questions out of 1,986 |
| LongMemEval-S cleaned | Evidence Hit@10 | 0.920000 (92.0%) | 500 conversations, 246,750 turns, 500 scored questions |
| PersonaMem-v1 32k | Multiple-choice accuracy | 0.679117 (67.9%) | 20 personas, 37 contexts, 6,425 visible turns; 400 correct out of 589 answers |

Evidence Hit@10 is the fraction of scored questions with at least one relevant
piece of evidence in the first 10 retrieval results. Four LoCoMo questions lack
evidence labels and are excluded from its retrieval-quality denominator; all
1,986 questions were included in the run.

PersonaMem measures four-choice answer accuracy, not evidence retrieval.
Its score cannot be ranked directly against the other two percentages.
These figures do not constitute a like-for-like comparison with Mem0.

## Conditions recorded in the historical report

- Source tree: clean; all three reported runs used the revision above.
- Hardware: x86_64 server, 32 CPUs, 125 GiB RAM.
- Embedding: Qwen/Qwen3-Embedding-4B; 2,560 dimensions; batch size 16.
- Vector index: HNSW. Retrieval: top 10. Workers: 8.
- Semantic candidate multiplier: 4; candidate limit: 1,024.
- Global dialogue-context budget: 29,400 bytes.
- PersonaMem answer model: gpt-5.6-terra.
- Runs were performed sequentially, without concurrent memory benchmarks.

## Completion and supporting measurements

The report records complete/passed results for all three datasets, with zero
projection-attempt and final-projection failures. PersonaMem records zero
generation failures, degraded retrieval queries and empty retrievals.

| Measurement | LoCoMo | LongMemEval-S cleaned | PersonaMem-v1 32k |
| --- | ---: | ---: | ---: |
| Projected memories | 5,882 | 246,001 | 6,421 |
| Mean retrieval/query latency | 86.904 ms | 691.069 ms | 208.229 ms |
| Evidence turn recall | 0.853641 | 0.784598 | Not applicable |
| MRR | 0.775251 | 0.831319 | Not applicable |

PersonaMem generated 589 answers. It recorded 4,348,066 input tokens
(including 2,259,968 cached-input tokens), 37,293 output tokens, 33,480 reasoning
tokens and 4,385,359 total tokens. Mean generation latency was 3,999.760 ms;
mean end-to-end latency was 4,207.989 ms. Token categories are preserved as
reported and should not be summed without checking their definitions.

## Dataset fingerprints

- LoCoMo: `79fa87e90f04081343b8c8debecb80a9a6842b76a7aa537dc9fdf651ea698ff4`
- LongMemEval-S cleaned: `d6f21ea9d60a0d56f34a05b609c79c88a451d2ae03597821ea3d5a9678c3a442`
- PersonaMem questions: `cccd34cf53e0bc4d9536c04cff5ca045156d9a4e227e83327112482840bbc93c`
- PersonaMem contexts: `217247ebfec9e8442fc53570c795ab69f21aad08745f7de78d9beab51b122d4a`

## Evidence limitations

The homepage uses the archived report's values. The repository's adjacent
`*-neural-full.json` files currently describe other revisions/runs and do not
reproduce this report's figures; they are not the source for the chart.
Per-question audit files were retained on the evaluation server and were not
committed with the report. This summary does not claim a fresh reproduction.