perfectrecall

# PerfectRecall

**Jev-powered memory for AI agents.** Try it on Hermes. Replace Mnemosyne while keeping your existing SQLite database, memory banks, and tool calls.

An independent 120-question LongMemEval-S experiment reduced final-answer errors from **62 to 17 (72.6%)**, using the same GPT-5.6 Luna caller with high reasoning effort. Accuracy rose from **48.3% to 85.8%**. These are results for a frozen predecessor of this release, not a fresh validation of the renamed package or an official leaderboard score. [Methods, results, regressions, and reproduction](benchmarks/README.md).

PerfectRecall asks Jev to evaluate every eligible memory against short criteria supplied by the calling agent. It returns the original evidence for that agent to reason over. There is no embedding model, vector index, or hidden query-writing model in the production package.

Jev is a remote decision model. It answers typed questions, for example whether one memory matches one criterion, and does not generate text. The default model is `typesafe/jev-1.13` on OpenRouter: you need an OpenRouter API key, and recall costs money.

## Install in Hermes

Hermes versions with Git plugin installation can install the repository as a directory plugin. Then run the native memory setup:

```sh
hermes plugins install https://github.com/arslanrakh/perfect-recall
hermes memory setup perfectrecall
```

This installs the repository's code and declared Python dependencies without a separate PerfectRecall pip installation. The setup command backs up configuration, selects PerfectRecall, and disables legacy source plugins while keeping their files. Set `OPENROUTER_API_KEY` in the Hermes environment and restart.

PerfectRecall is not in the Hermes plugin catalog, so `hermes plugins install perfectrecall` by bare