PerfectRecall

It is a memory provider: Hermes calls it before and after turns, so prefetch injects relevant memory into context and post-turn synchronization saves useful user messages. Jev evaluates every eligible memory against short criteria supplied by the calling agent and returns the original evidence for that agent to reason over.

On 120 independent LongMemEval-S questions, the README reports final-answer errors falling from 62 to 17 (72.6%) and accuracy rising from 48.3% to 85.8%, using the same GPT-5.6 Luna caller with high reasoning effort. Those results are for a frozen predecessor of this release, not a fresh validation or an official leaderboard score.

For Hermes versions with Git plugin installation, run hermes plugins install https://github.com/arslanr-com/perfectrecall followed by hermes memory setup perfectrecall. Alternatively install the Python package into the same Python environment that runs Hermes (Python 3.10 or newer), then run python -m perfectrecall.install.

Set OPENROUTER_API_KEY in the environment used to start Hermes, or in your existing Hermes secret configuration, then restart Hermes. The jev-status command reports configuration and key presence; it does not print the key or make an API call. Hermes itself must already be installed.

No. The README states there is no embedding model, vector index, or hidden query-writing model in the production package: Jev evaluates each eligible memory against the calling agent's criteria, and PerfectRecall returns the original evidence rather than similarity hits.

Yes. Setting default_scope: global under memory.perfectrecall deliberately shares memories across sessions within the same bank; omit it to preserve the inherited session scope. sync_roles controls which roles are synchronized, and assistant messages are excluded by default.

The headline accuracy numbers describe a frozen predecessor, not the renamed package, and the README says a PyPI release is not assumed, so the repository is the installation source. Economy prefetch may reuse the previous search question; explicit recall always searches the supplied question.