claw-self-improving-plus（爪自强加）

Claw Self Improving Plus

Build a conservative learning pipeline. Optimize for signal, not clutter.

Core stance

Do not auto-rewrite long-term memory or behavior files by default.

Use this flow:

1. Capture raw learning candidates.
2. Normalize them into a structured schema.
3. Score each item for promotion value.
4. Detect duplicates or merge candidates.
5. Consolidate repeated learnings into stronger records.
6. Build a prioritized learning backlog.
7. Draft anchored candidate patches.
8. Review patches with human approval.
9. Apply only approved patches.

Learning types

Use these types:

• mistake: the agent did something wrong
• correction: the user corrected a wrong assumption or behavior
• discovery: a useful fact about environment, tools, preferences, or workflow
• decision: a durable preference, policy, or chosen design
• regression: a known failure mode that should not recur

Minimal record schema

Store each learning candidate as JSON with these fields:

• id: stable slug or timestamped id
• timestamp
• source
• type
• summary
• details
• evidence
• confidence
• reuse_value
• impact_scope
• promotion_target_candidates
• status
• related_ids

Default enums:

• confidence: low|medium|high
• reuse_value: low|medium|high
• impact_scope: single-task|project|workspace|cross-session
• status: captured|scored|merged|promoted|rejected

Routing rules

Promote by destination, not vibes:

• SOUL.md: durable style, personality, voice rules
• AGENTS.md: operating rules, workflows, safety/process lessons
• TOOLS.md: environment-specific commands, paths, model/tool preferences
• MEMORY.md: important long-term facts about user, projects, decisions, history
• daily/raw store only: low-confidence or highly local observations

If a learning does not clearly deserve promotion, keep it in the raw log.

Scoring heuristic

Score each record on five dimensions:

1. reuse_value: will this help again?
2. confidence: how well supported is it?
3. impact_scope: how broadly does it matter?
4. promotion_worthiness: should it become a lasting rule or memory?
5. promotion_target_candidates: where should it go if promoted?

Use this practical rubric:

• High promotion priority: repeated mistake, explicit user preference, environment fact that breaks tasks, regression with real cost
• Medium priority: useful workflow pattern seen more than once
• Low priority: one-off trivia, speculative interpretation, emotional noise, temporary state

Anchored patch generation

Prefer anchored insertion or exact replacement over blind append.

Each patch may contain:

• target_file
• anchor
• insert_mode
• old_text
• new_text
• suggested_entry
• approved
• review_status

Use exact replacement when the old text is known. Use anchored insertion when the destination section is known. Use append only as fallback.

Learning store layout

Use a stable .learnings/ structure. See references/learning-store-layout.md.

Recommended files:

• .learnings/inbox.jsonl
• .learnings/scored.jsonl
• .learnings/merge.json
• .learnings/patches.json
• .learnings/apply-report.json
• .learnings/archive/

Default workflow

1. Capture

Append raw learnings into .learnings/inbox.jsonl.

Use scripts/capture_learning.py to create normalized records.

2. Score

Run scripts/score_learnings.py on the inbox or a batch export.

3. Review duplicates

Run scripts/merge_candidates.py to group likely duplicates.

4. Draft patches

Run scripts/draft_patches.py to produce anchored reviewable patch candidates.

5. Review

Use scripts/review_patches.py to list, approve, reject, or skip candidates.

Examples:

python scripts/review_patches.py .learnings/patches.json list
python scripts/review_patches.py .learnings/patches.json act --index 1 --action approve
python scripts/review_patches.py .learnings/patches.json act --index 2 --action reject --note "too vague"

6. Apply only after approval

Run scripts/apply_approved_patches.py.

This script only applies entries explicitly approved. It validates allowed targets, supports --dry-run, skips duplicate entries already present, and prefers exact replacement, then anchored insertion, then append fallback.

Output style

When reporting results, use this structure:

• new_candidates: count
• high_priority: count
• merge_groups: count
• patch_candidates: short bullet list
• needs_human_review: yes

Resources

References

• Scoring rubric: see references/scoring-rubric.md
• Patch target guide: see references/promotion-targets.md
• Learning store layout: see references/learning-store-layout.md

Scripts

• scripts/capture_learning.py
• scripts/score_learnings.py
• scripts/merge_candidates.py
• scripts/draft_patches.py
• scripts/detect_patch_conflicts.py
• scripts/consolidate_learnings.py
• scripts/build_backlog.py
• scripts/age_backlog.py
• scripts/review_backlog.py
• scripts/check_existing_promotions.py
• scripts/review_patches.py
• scripts/render_review.py
• scripts/apply_approved_patches.py
• scripts/archive_batch.py
• scripts/run_pipeline.py