weed-memories

name	weed-memories
description	Scan auto-memory files across all projects on this machine and identify entries that contradict the current project's CLAUDE.md direction, then offer to delete the biasing ones. Use when the auto-memory system is surfacing stale framing that anchors thinking toward archived or retired work.
disable-model-invocation	true

Weed biasing memories

Find memory files — across every project on this machine — that would bias a future session toward a direction the user has since rejected, and offer to delete them. Never delete without explicit user approval.

Process

1. Establish ground truth for the current project.

   • Read the current project's CLAUDE.md at the repo root.
   • If .claude-memory/MEMORY.md exists in the repo, read it — it is authoritative over the auto-memory.
   • Extract the active direction, scope, and any explicit "off-scope" / "archived" markers.

2. Enumerate every memory file on this machine.

   • Use Glob: /home/hammar/.claude/projects/*/memory/*.md
   • Include all projects, not only the current one. Cross-project anchoring is the bug we're hunting.

3. Classify each memory file by reading it:

   • keep — consistent with the active direction; or purely general process / tooling guidance (workflow preferences, planning rules, ClickHouse gotchas, language/perf findings, data-infrastructure facts).
   • retire-candidate — contradicts the active direction; describes archived work; frames the task in a way the user has explicitly rejected; references specific tables, files, or artifacts that no longer exist.
   • verify — references specific code, paths, or state that cannot be confirmed without reading the repo. Not auto-retired; just flagged.

4. Produce a report grouped by classification. For each retire-candidate, show:

   • Full path to the memory file
   • One-line reason (e.g. "frames project as momentum-prediction; current direction is neutral feature extraction")
   • First 3–5 lines of the memory body for user context

5. Ask before deleting. List the retire candidates and wait for explicit approval (per-file or a batch "delete all listed"). After approval:

   • Delete the file
   • Remove its line from the containing directory's MEMORY.md index
   • Do not delete the MEMORY.md index itself.

What counts as biasing

• Describes a task/direction the user has since archived (e.g. spike prediction when current work is neutral feature extraction)
• Names tables, files, or artifacts that were dropped / archived / moved
• Locks in a framing ("two-tier momentum strategy", "predict pump events", "ride the retail tail") that contradicts the current CLAUDE.md
• Recommends a specific detector or algorithm whose scaffolding has been archived

What is NOT biasing

• General ClickHouse / shell / language gotchas
• Process preferences (save queries as script files, planning workflow, thoroughness, no-autostart, detection-scope discipline, no calendar analysis)
• Language or performance findings that are codebase-agnostic
• Data-infrastructure descriptions (data gaps, import patterns) — unless CLAUDE.md has superseded them

Caveats

• Do not extract zip archives, even if they are referenced in memories.
• Do not read files inside archive/ directories of any project.
• Do not touch .claude-memory/ directories inside any repo — those are user-authoritative and out of scope.
• When borderline, classify verify rather than retire-candidate. Err toward preservation and let the user decide.
• Run this skill /weed-memories manually only; do not auto-invoke, side effects are destructive.