jevmem is a memory tool that saves decisions, rules, and failed attempts made in Claude Code and Cursor sessions to JEVMEM.md, so that they can be loaded back in the next conversation. The 2026-10-19 commit 981c6f3 reflects the release gate result for 0.7.6 in the documentation. With Claude Code 2.1.303, 17 scenarios were run 3 times each, and all 51 cases in total passed.
The gate was executed with the command scripts/e2e.sh --runs 3 --scenario full, on the final build 6d3cf43 with the real Claude Code and the real Jev attached, from 13:47Z to 14:32Z. The model was not pinned: claude-fable-5-1 handled 156 of 177 runs, and claude-sonnet-5-5 handled 21. The results document contains the 51 passes along with 57 PASS lines, because the nokey scenario prints one line per form. There were no hook errors and no timeouts. The same 0.7.6 record also shows, in a 66-turn benchmark, the auto mode recording save/skip 65/66, 98.5%, and by kind 63/66, 95.5%, leaving p50 238ms and $0.000179 per decision; on the Cursor side, with agent CLI 2026.10.16, the cursor-plugin scenario passed 3 out of 3, and 9 agent executions passed.
However, in 0.7.6 the Cursor cursor scenario could not be run. The retry was also blocked by the monthly additional usage limit of the Cursor plan and ended with 0 cases before any model request, and the rule cost measured by Cursor's own criteria remains an estimate in cursor-rule-cost-2026-10-19.json. The build tree scan was 97/100 with plugin-scanner 3.0.123 and 100/100 with 3.16.5, and there was no high verdict. The commit modifies CHANGELOG.md, results/README.md, and results/CURRENT.json, and newly adds the 1914-line log results/e2e-2026-10-19-v076-full.txt (1918 insertions, 1 deletion). In other words, the Claude Code path was confirmed on the final build, but the Cursor-side cost is still an estimate, so it is too early to settle the Cursor configuration on the basis of this commit alone.