mnesis

<p align="center">
  <img src="https://raw.githubusercontent.com/Lucenor/mnesis/main/docs/images/logo_icon.png" alt="mnesis logo" width="120"><br><br>
  <img src="https://raw.githubusercontent.com/Lucenor/mnesis/main/docs/images/logo_wordmark.png" alt="mnesis" width="320"><br><br>
  <em>Lossless Context Management for long-horizon LLM agents</em>
  <br><br>
  <a href="https://pypi.org/project/mnesis/"><img src="https://img.shields.io/pypi/v/mnesis?color=5c6bc0&labelColor=1a1a2e" alt="PyPI"></a>
  <a href="https://pypi.org/project/mnesis/"><img src="https://img.shields.io/pypi/pyversions/mnesis?color=5c6bc0&labelColor=1a1a2e" alt="Python"></a>
  <a href="LICENSE"><img src="https://img.shields.io/badge/license-Apache%202.0-5c6bc0?labelColor=1a1a2e" alt="License"></a>
  <a href="https://github.com/Lucenor/mnesis/actions/workflows/ci.yml"><img src="https://img.shields.io/github/actions/workflow/status/Lucenor/mnesis/ci.yml?color=5c6bc0&labelColor=1a1a2e" alt="CI"></a>
  <a href="https://codecov.io/github/Lucenor/mnesis"><img src="https://img.shields.io/codecov/c/github/Lucenor/mnesis?color=5c6bc0&labelColor=1a1a2e" alt="Coverage"></a>
  <a href="https://mnesis.lucenor.tech"><img src="https://img.shields.io/badge/docs-mnesis.lucenor.tech-5c6bc0?labelColor=1a1a2e" alt="Docs"></a>
  <a href="https://github.com/Lucenor/mnesis/attestations"><img src="https://img.shields.io/badge/provenance-attested-5c6bc0?labelColor=1a1a2e&logo=githubactions&logoColor=white" alt="Attestation"></a>
  <a href="https://scorecard.dev/viewer/?uri=github.com/Lucenor/mnesis"><img src="https://api.scorecard.dev/projects/github.com/Lucenor/mnesis/badge" alt="OpenSSF Scorecard"></a>
</p>

---

LLMs suffer from **context rot**: accuracy degrades 30–40% before hitting nominal token limits — not because the model runs out of space, but because reasoning quality collapses as the window fills with stale content.

The standard fix — telling the model to "summarize itself" — is unreliable. The model may sile