# How we got 99.2% recall@5 on LongMemEval-S (and published the harness) — Mnemo Memory

> 496 out of 500, no reranker, no LLM in the retrieval loop — and a depth sweep that shows it isn't a candidate-pool trick. The number, the method, and why we published the harness instead of just the headline.

Canonical URL: https://mnemohq.com/blog/how-we-got-99-2-recall-at-5

496 out of 500, no reranker, no LLM in the retrieval loop — and a depth sweep that shows it isn't a candidate-pool trick. The number, the method, and why we published the harness instead of just the headline.

## About Mnemo

Mnemo is long-term memory infrastructure for AI agents. Write facts through the API, retrieve them in natural language across seven fused retrieval signals, and get every answer back with the source attached. US-hosted, model-agnostic, with a free tier of 1K writes and 10K searches per month.

## Explore Mnemo

- [Mnemo API documentation](https://mnemohq.com/docs)
- [Pricing](https://mnemohq.com/pricing)
- [Benchmarks](https://mnemohq.com/benchmarks)
- [Company Brain for Slack](https://mnemohq.com/brain)
- [Ask Mnemo widget](https://mnemohq.com/widget)
- [Data-source connectors](https://mnemohq.com/connectors)
- [Framework integrations](https://mnemohq.com/integrations)
- [Trust and security](https://mnemohq.com/trust)
- [Contact](https://mnemohq.com/contact)
- [llms.txt](https://mnemohq.com/llms.txt)
- [OpenAPI 3.1 specification](https://mnemohq.com/openapi.json)

---

This is the Markdown representation of https://mnemohq.com/blog/how-we-got-99-2-recall-at-5, served by content negotiation on `Accept: text/markdown`. Machine-readable index: https://mnemohq.com/llms.txt · https://mnemohq.com/openapi.json · https://mnemohq.com/sitemap.xml