What it is
Keeps long agent sessions from re-sending stale tool output.
What it does: Classifies each span as keep, compress or archive, with Jev for the hard cases.
Why it belongs in the showcase: Rules first, Jev second, and a way to measure false evictions.
Caveats: Token savings, not proven cost savings.
Evidence
Public MIT repo (4 stars as of 2026-10-21) with a documentation site. The README shows a live-API demo: for an invoice reconciliation task, two shipment-delay spans got the same keyword-overlap score (0.11), while Jev scored the relevant one 0.75 and the unrelated one 0.45. Evicted spans keep a keyword index and can be rehydrated from snapshots; an analyze_regret pass compares a no-eviction baseline run with the collected run. Integrations for LangGraph and Strands are optional extras.
Caveats
- Reports token counts, not prompt-cache effects, so it does not show money saved
- The regret check is keyword-overlap evidence, not proof of cause
- Needs your own Jev API key
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.