Release Notes
Changed
memory.md carries only facts no repo file holds. The Latest Project State
and Next Actions sections are removed: they restated tasks/todo.md,
docs/handoffs/, and git, which agents read directly. Output is now a scope
line, Memory health, and the fact libraries.- Takeaway selection scans the store instead of running four fixed queries.
Candidates previously had to resemble one of those queries to be scored at
all, so high-priority chunks could go unranked indefinitely. search_score
and multi_query are removed from the priority breakdown.
- Verified retrieval outcomes now feed ranking.
utility was hardcoded to
0.0; it is derived from outcome_priors and capped at decision tier.
Added
- Repo-novelty gate: takeaways whose distinctive tokens are already contained
in tasks/**, docs/handoffs/, README.md, CLAUDE.md, or AGENTS.md are
suppressed with reason covered_by_repo:<path>.
memd outcome-scan: derives verified retrieval outcomes from agent session
transcripts by matching distinctive literals from served chunks against
commands run afterwards. Reads tool-call inputs only, so being shown a chunk
again is never counted as using it. Supports --dry-run.memd search accepts a positional query, --limit as an alias for --k,
and memd get accepts a positional chunk id. The previous forms still work
and supplying both errors.[[server.project_aliases]] entries make chunks stored under a historical
tenant readable from the canonical one without moving data, preserving
chunk_id and timestamp_created.
Install memd 1.7.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/fmschulz/memd/releases/download/v1.7.0/memd-installer.sh | sh
Download memd 1.7.0
