A human example
A long debugging session contains repeated file reads alongside one failed attempt that must not be forgotten.
What the caller supplies
The caller supplies bounded records and the current goal while preserving the original session.
What happens next
Jev recommends keep or review. The agent can recover all records and retains instructions through deterministic rules.
Illustrative example, not a recorded result.
Potential value: high
Context pressure is common for frequent coding-agent users. A reliable selector could make long tasks easier to continue.
Evidence confidence: low
Firsthand Reddit enthusiasm and conflicting X previews establish interest, not successful compaction or lower total cost.
The rating describes support for this claim. It is separate from Jev's returned probability. How we assign ratings.
Evidence, including disagreement
- Model routing: test whether Jev adds value. reported. No-Jev ablation matched the hybrid result.
- A user tries Jev to reduce a long agent history. anecdotal. A user found history selection useful; quality and cache costs need a test.
- X posts disagree about compaction by filtering. anecdotal. Conflicting firsthand views identify a question to test, not a settled result.
The next test
20 consented or purpose-authored long sessions, each with later questions about decisions, constraints and failed attempts.
Compare against
- Existing agent compaction
- Deterministic deduplication
- No selection
Measure
- Later task success
- Lost constraints
- Repeated failed actions
- Cache rewrite cost
- Recovery reads
Decision after the test
Do not automate destructive history edits. Retain the experiment only if task continuity matches the baseline and total cost or time improves.
The report will retain inputs, question versions, every attempt and failure examples. We will update the confidence rating after reviewing the result.
Use a related Engine recipe
Recipes are implementation starting points. Their presence does not mean the protocol above has passed.