What's new
ai-memory link — connect memories to git commits
Scan recent commits and auto-link them to the memories they implement. Uses weighted Jaccard scoring with a three-band threshold model (auto-link ≥0.70 / suggest 0.40–0.70 / drop <0.40). Links stored as invisible <!--links--> comment blocks in memory frontmatter. Supports --dry-run, --clear-auto, --since, --auto-threshold. Dashboard graph now shows green "co-implementing" edges between memories that share a commit.
ai-memory init --schedule — daily auto-extraction
Register a daily extract --incremental job with the OS-native scheduler: launchd (macOS), crontab (Linux), Task Scheduler (Windows). Runs at 09:00 local time. Remove with --unschedule. No new runtime dependencies.
Post-extraction dedup improvements (F1 64.1% → 75%+ target)
- Single-chunk dedup fix — single-LLM-call extractions (the common case) now go through
deduplicateMemories; previously only multi-chunk runs were deduplicated. - Cross-type TODO subsumption — TODOs whose content is ≥75% covered by a same-extraction decision/architecture/convention are automatically dropped as sub-steps.
Dashboard graph enhancements
Edge type differentiation (indigo solid = same conversation / gray dashed = shared keyword / green arrow = co-implementing); type filter toggle buttons; hover highlight; stats bar with node/edge counts.