Skill memory reflect

Skill fredrick84823/fstack/skills/skill-memory-reflect

Reflect on /improve Skill Evolution Memory. Use when the user asks to process accumulated improve signals, deduplicate skill gaps, generate or refresh memory/claims/{skill}.md, generate memory/eval-cases/{skill}.json, enrich old memory records, or produce a prioritized /improve worklist from signals.jsonl.From its SKILL.md

Install

npx -y skills add fredrick84823/fstack --skill skill-memory-reflect

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.3 KB, as published. Nobody here has run it

Skill Memory Reflect

Turn raw /improve memory events into reviewable skill-evolution artifacts. Use this skill after hooks or migrations have written memory/signals.jsonl and the user wants recurring gaps consolidated into concrete claims, eval cases, and implementation priorities.

Workflow

Locate the improve memory directory. Prefer an explicit --memory-dir; otherwise use /Users/fredrick/.agents/skills/improve/memory for installed user memory, or skills/improve/memory when working inside a repo fixture.
Run a dry reflection package first:

python3 skills/skill-memory-reflect/scripts/reflect_memory.py --memory-dir /Users/fredrick/.agents/skills/improve/memory --dry-run

Review the summary. Check especially groups whose affected_rule or gap_type was inferred from low-confidence text. If the result is noisy, narrow with --target-skill <skill>.
Write artifacts only after the dry run looks coherent:

python3 skills/skill-memory-reflect/scripts/reflect_memory.py --memory-dir /Users/fredrick/.agents/skills/improve/memory --write

Use the generated worklist to drive /improve changes. Do not automatically edit target skills from this reflection alone; treat claims and eval cases as a review package.

Outputs

The script writes derived artifacts under the selected memory directory:

claims/{skill}.md: deduplicated candidate claims with evidence, inferred expected and actual behavior, risk notes, and proposed rule updates.
eval-cases/{skill}.json: A/B/C/D eval cases grounded in signal evidence. A and B cover recall/regression; C covers downstream or workflow impact; D is a false-positive trap.
worklists/YYYYMMDDTHHMMSSZ-skill-memory-reflect.md: prioritized /improve worklist across skills.
skill-graph.json: refreshed claim and eval-case indexes.

Reflection Rules

Preserve signals.jsonl as the raw source of truth. Do not delete or rewrite raw records.
Group signals by target skill, inferred affected rule, inferred gap type, and normalized gap text. Use repeated evidence to prioritize, but keep single high-risk signals visible.
Prefer explicit fields from memory records. Infer missing affected_rule, gap_type, expected_behavior, and actual_behavior from gap text only when the original record has unknown or empty values.
Flag low-confidence groups instead of hiding them. Many backfilled signals are useful even when old records lack structured fields.
Generate eval cases that test both sides of the rule: a positive recall case and a clean/trap case that must not emit a new GAP.
Leave queue status changes to /improve or a human review step. Reflection produces derived recommendations; it does not mark pending signals resolved.

Script Notes

scripts/reflect_memory.py is deterministic and does not call external services. It uses heuristic text normalization so an agent can perform the semantic review around it instead of reimplementing parsing and file writing each time.