ReflexLearn

The OpenClaw skill that turns silence into signal.
No ratings. No feedback buttons. Just pure implicit learning from how users actually behave.

---

What It Does

Every time a user repeats a similar question, ReflexLearn infers the previous answer failed — and writes a structured reflection to MEMORY.md. Every time a question is not repeated, it writes a positive reinforcement entry. No explicit feedback required. Ever.

v1.1.0 adds critical safeguards that make this production-ready:

• Multi-repeat threshold: Requires 2+ repeats before flagging failure (avoids false positives from natural follow-ups)
• Modifier detection: "Write a Python script to parse JSON with type hints" is treated as preference extraction, not failure
• Tiered memory: In cautious mode (default), SOUL.md updates are staged in reflexlearn-pending.md for your review before being applied
• Aggressive mode: For power users who want instant SOUL.md writes on first confirmed failure
• Slash commands: /reflex status and /reflex ignore-last for manual control

---

Before / After

Before ReflexLearn:

User: "Explain async/await in Python"
Agent: [gives a wall of text with no examples]
User: "Explain async/await in Python"   ← agent has no idea this was bad
Agent: [same wall of text again]

After ReflexLearn:

User: "Explain async/await in Python"
Agent: [gives a wall of text with no examples]
→ ReflexLearn logs interaction, signal=neutral

User: "Explain async/await in Python"
→ ReflexLearn: Repeat #1/2 (sim=0.9821) — watching

User: "Explain async/await in Python"
→ ReflexLearn: *** FAILURE SIGNAL *** Repeat threshold reached
→ MEMORY.md: "User repeated 'Explain async/await' 2+ times — previous answer insufficient."
→ reflexlearn-pending.md: "Proposed: improve:explain-async-await — add examples and clarity"

Agent (next session): [reads MEMORY.md, gives concise answer with code examples]
User: satisfied — doesn't repeat the question
→ ReflexLearn heartbeat: Positive reinforcement written
→ SOUL.md: "keep:explain-async-await — user satisfied, maintain this style"

Modifier detection (intentional refinement, NOT flagged as failure):

User: "Write a Python script to parse JSON"
Agent: [writes script without type hints]
User: "Write a Python script to parse JSON with type hints"
→ ReflexLearn: Modifier detected ("with type hint") → preference extraction
→ MEMORY.md: "User prefers Python scripts with type hints"
→ reflexlearn-pending.md: "Proposed: pref:with-type-hint — user prefers type hints"

---

Installation

Just add to soul.md:

## Skills
- reflex-learn

Manual install:

# 1. Copy skill folder
mkdir -p ~/.openclaw/skills/reflex-learn
cp -r /path/to/ReflexLearn/* ~/.openclaw/skills/reflex-learn/

# 2. Install dependencies
pip install -r ~/.openclaw/skills/reflex-learn/requirements.txt

# 3. Test it
python3 reflex_learn.py --query "Hello world"
python3 reflex_learn.py --query "Hello world"
python3 reflex_learn.py --query "Hello world"
# → Triggers FAILURE SIGNAL on 3rd run (repeat_count=2 >= threshold=2)

# 4. Check results
cat ~/.openclaw/MEMORY.md
cat ~/.openclaw/reflexlearn-pending.md

---

Configuration

Open SKILL.md and edit the config block:

SIMILARITY_THRESHOLD: 0.85       # How similar = "same question" (0.0–1.0)
LOOKBACK_INTERACTIONS: 10        # How many past queries to check
POSITIVE_REINFORCEMENT_DELAY: 3  # Interactions to wait before confirming success
REPEAT_COUNT_THRESHOLD: 2        # Repeats needed before flagging failure
SESSION_WINDOW_MINUTES: 60       # Time window for counting repeats
MODE: cautious                   # "cautious" (safe) or "aggressive" (instant writes)

Cautious mode (default): Proposed SOUL.md updates are written to ~/.openclaw/reflexlearn-pending.md. Review and approve them by moving entries to SOUL.md.

Aggressive mode: Updates are written directly to SOUL.md on first confirmed failure. Best for users who trust the signal and want fully automated learning.

---

Slash Commands

Command	Effect
/reflex status	Show interaction count, current config, pending review count
/reflex ignore-last	Remove the last logged interaction (useful after test queries)

---

Signal Reference

Signal	Trigger	Action
neutral	No similar query in history	Log interaction only
watching	Similar query, below repeat threshold	Log with watching signal
preference	Similar query + modifier words detected	Extract preference → MEMORY.md + pending
negative	Repeat threshold reached	Reflection → MEMORY.md + pending (or SOUL.md in aggressive)
reinforced	Query not repeated in next N interactions	Reinforcement → MEMORY.md + SOUL.md

---

Optional: Ollama Integration

For richer, AI-generated reflections (requires Ollama running locally):

python3 reflex_learn.py --query "your query" --use-ollama --ollama-model llama3

If Ollama is unavailable, ReflexLearn falls back to a structured local reflection automatically.

---

File Reference

File	Purpose
~/.openclaw/MEMORY.md	All reflections and reinforcements (always written)
~/.openclaw/SOUL.md	Learned patterns (aggressive mode or after manual approval)
~/.openclaw/reflexlearn-pending.md	Staged updates awaiting review (cautious mode)
~/.openclaw/reflex_history.json	Interaction history with embeddings (internal)

---

Submitting to ClawHub

1. Push this folder to a public GitHub repo: skills/<your-username>/reflex-learn/
2. Submit a PR to openclaw/skills
3. Add an entry to awesome-openclaw-skills:

   - [reflex-learn](link) - Turns query repetition into implicit feedback for continuous learning.

---

License

MIT — see LICENSE.