Self-Correction Analysis — Status & Implementation Plan

Issue: Feature: Nightly Self-Correction Analysis — Automated Failure Detection & Remediation #34

Last updated: 2026-02-18

---

Current status: Implemented (multi-language)

The pipeline is implemented in the memory-hybrid extension with multi-language support via .language-keywords.json (see openclaw hybrid-mem build-languages).

Component	Status
Phase 1: Extract	✅ openclaw hybrid-mem self-correction-extract — uses merged correction signals (English + translated from .language-keywords.json)
Phase 2: Analyse	✅ Prompt self-correction-analyze.txt; LLM via self-correction-run
Phase 3: Remediate	✅ MEMORY_STORE + TOOLS.md append (cap 5); AGENTS/SKILL as proposals only
Phase 4: Report	✅ memory/reports/self-correction-YYYY-MM-DD.md
Protocol doc	✅ SELF-CORRECTION-PIPELINE.md
Cron job	Optional — add to OpenClaw jobs; see pipeline doc
Tests	✅ Extract (self-correction-extract.test.ts), TOOLS section (tools-md-section.test.ts), config (config.test.ts for selfCorrection)
Semantic dedup	✅ Optional (default on): skip MEMORY_STORE when embedding similarity ≥ threshold (config selfCorrection.semanticDedup, semanticDedupThreshold).
TOOLS sectioning	✅ Rules inserted under configurable section (e.g. “Self-correction rules”), not appended at end. Dedup within section.
Auto-rewrite vs approve	✅ Default: suggest TOOLS rules in report; apply with --approve. Opt-in autoRewriteTools: true: LLM rewrites TOOLS.md to integrate new rules (no duplicates/contradictions).
Phase 2 via spawn	✅ Optional: analyzeViaSpawn: true + spawnThreshold → Phase 2 runs via openclaw sessions spawn (e.g. Gemini) for large batches.

Multi-language: Run openclaw hybrid-mem build-languages once (or when you add languages). Correction phrases are translated into your top languages so detection works in all of them.

---

Implementation checklist (done)

1. Phase 1 — Extract ✅
   Implemented in services/self-correction-extract.ts; CLI self-correction-extract. Uses getCorrectionSignalRegex() (English + .language-keywords.json translations). Skip filters applied.

2. Phase 2 — Analyse ✅
   Prompt prompts/self-correction-analyze.txt; invoked inside self-correction-run with configurable model.

3. Phase 3 — Remediate ✅
   MEMORY_STORE (with dedup), TOOLS_RULE (append-only), AGENTS_RULE/SKILL_UPDATE as proposals; cap 5.

4. Phase 4 — Report ✅
   Written to memory/reports/self-correction-YYYY-MM-DD.md.

5. Documentation ✅
   SELF-CORRECTION-PIPELINE.md — protocol and multi-language usage. Cron job is optional (user/add-on).

6. Tests ✅
   self-correction-extract.test.ts — extraction and skip filters.

---

Success criteria (from issue)

• Detects ≥80% of user corrections within 24 hours.
• Auto-remediates ≥50% of detected issues (memory stores + TOOLS rules).
• Reduces repeat corrections by ≥60% over 30 days.
• False positive rate <20%.

---

References

• GitHub issue #34: Nightly Self-Correction Analysis
• Related: Procedural Memory #23 (stores learned procedures; this pipeline discovers what to learn).
• Manual analysis: memory/reports/failure-analysis-v2-2026-02-18.md (if present in workspace).
• Raw extracts POC: memory/reports/correction-extracts-raw.md (if present).