true-recall-out

name: true-recall-out description: | Smart memory curation system for OpenClaw. Stages all conversations in Redis, then uses AI to evaluate and only persist "things of note" to Qdrant long-term memory.

Use when: User wants intelligent memory filtering, not indiscriminate storage. Trigger phrases: "smart memory", "curate memories", "filter what to remember", "true recall", "tr-out", "review today's memories". metadata: openclaw: os: ["linux"]

True-Recall-Out: Smart Memory Curation

Philosophy

Don't remember everything. Remember what matters.

Instead of dumping every exchange into permanent storage, this system:

1. Stages everything in Redis (24-48 hour buffer)
2. Curates using AI judgment — only "things of note" graduate to Qdrant
3. Preserves full context for what does matter

Architecture

Conversation Flow:
┌─────────────┐     ┌──────────────┐     ┌─────────────────┐
│   Rob asks  │────▶│ Redis Buffer │────▶│ AI Curation Job │
│  AI_NAME replies│     │  (24-48hrs)  │     │  (Daily/On-demand)│
└─────────────┘     └──────────────┘     └─────────────────┘
                                                  │
                    ┌─────────────────┐          │
                    │  Qdrant Long-term │◀─────────┘
                    │    Memory         │
                    │  (Things of Note) │
                    └─────────────────┘

Curation Criteria

The AI curator evaluates each conversation against these criteria:

STORE (High Value):

• Decisions made or conclusions reached
• Technical solutions that worked
• Preferences stated ("I like...", "I prefer...", "Always...")
• Project plans, architecture decisions
• Important context about people, relationships
• Rules or constraints established
• Things you explicitly said to remember

DISCARD (Low Value):

• Casual banter, greetings, goodbyes
• Debugging back-and-forth that didn't resolve
• Repetitive questions with same answers
• Transient status checks ("what's in redis?")
• Test messages, experiments
• "Thanks", "Got it", "Cool" exchanges

Scripts

1. stage_to_redis.py — Automatic Staging

Stages every conversation turn to Redis buffer.

# Auto-called after each exchange
python3 scripts/stage_to_redis.py \
  --user-id "USER_ID" \
  --user-msg "What about the server?" \
  --ai-response "The server is at localhost" \
  --turn 1

2. curate_memories.py — AI Curation

Reviews staged conversations and promotes "things of note" to Qdrant.

# Review last 24 hours
python3 scripts/curate_memories.py --user-id "USER_ID" --hours 24

# Dry run (see what would be stored without storing)
python3 scripts/curate_memories.py --user-id "USER_ID" --hours 24 --dry-run

# Review specific date range
python3 scripts/curate_memories.py --user-id "USER_ID" --date 2026-02-21

3. search_memories.py — Retrieval

Search curated memories (replaces generic qdrant-memory search).

# Quick search
python3 scripts/search_memories.py "server IP"

# With context
python3 scripts/search_memories.py "server" --with-context

# Filter by importance
python3 scripts/search_memories.py "project" --importance high

Usage Flow

Daily Workflow

1. Chat normally throughout the day
   ↓ (auto-stage to Redis)
   
2. End of day / Morning review
   "curate my memories" or cron runs
   ↓ (AI reviews, judges, stores "things of note")
   
3. Qdrant now has only valuable memories
   ↓
   
4. Future conversations retrieve quality context

Manual Commands

Configuration

Edit config.yaml:

curation:
  # Model for judging importance
  model: "qwen3:4b-instruct"
  ollama_url: "http://localhost:11434"
  
  # Criteria weights
  weights:
    decision_made: 1.0
    technical_solution: 0.9
    preference_stated: 0.9
    project_plan: 0.8
    personal_context: 0.7
    rule_established: 0.9
    explicit_remember: 1.0
    
  # Threshold for storage
  store_threshold: 0.6
  
  # Auto-curation schedule
  auto_curation:
    enabled: true
    cron: "0 3 * * *"  # 3 AM daily
    
redis:
  host: "localhost"
  port: 6379
  buffer_key: "tr-out:buffer:{user_id}"
  retention_hours: 48
  
qdrant:
  url: "http://localhost:6333"
  collection: "AI_NAME_memories"

Integration with OpenClaw

Option 1: Auto-stage all exchanges (Recommended)

Add to OpenClaw config to auto-call stage_to_redis.py after each exchange:

{
  "memory": {
    "mode": "true-recall-out",
    "stage_script": "/path/to/true-recall/tr-out/true-recall-out/scripts/stage_to_redis.py",
    "auto_stage": true
  }
}

Option 2: Manual staging

Don't auto-stage. User says "stage this" when they want something buffered.

Option 3: Hybrid (Smart Auto)

Auto-stage, but use lightweight classifier to skip obvious low-value exchanges (greetings, "thanks", etc.) before hitting Redis.

Comparison: Old vs True-Recall-Out

Status

• ✅ Concept & architecture
• ✅ SKILL.md
• ✅ Config spec
• ⏳ Scripts (stage_to_redis.py, curate_memories.py, search_memories.py)
• ⏳ Testing
• ⏳ OpenClaw integration

Created: 2026-02-21 Purpose: Remember what matters, forget the noise.