Total Recall

Total Recall — Autonomous Agent Memory

The only memory skill that watches on its own.

No database. No vectors. No manual saves. Just an LLM observer that compresses your conversations into prioritised notes, consolidates when they grow, and recovers anything missed. Five layers of redundancy, zero maintenance. ~$0.00/month (using free-tier models).

While other memory skills ask you to remember to remember, this one just pays attention.

Architecture

Layer 1: Observer (cron, every 15-30 min)
    ↓ compresses recent messages → observations.md
Layer 2: Reflector (auto-triggered when observations > 8000 words)
    ↓ consolidates, removes superseded info → 40-60% reduction
Layer 3: Session Recovery (runs on every /new or /reset)
    ↓ catches any session the Observer missed
Layer 4: Reactive Watcher (inotify daemon, Linux only)
    ↓ triggers Observer after 40+ new JSONL writes, 5-min cooldown
Layer 5: Pre-compaction hook (memoryFlush)
    ↓ emergency capture before OpenClaw compacts context

What It Does

• Observer reads recent session transcripts (JSONL), sends them to an LLM, and appends compressed observations to observations.md with priority levels (high, medium, low)
• Reflector kicks in when observations grow too large, consolidating related items and dropping stale low-priority entries
• Session Recovery runs at session start, checks if the previous session was captured, and does an emergency observation if not
• Reactive Watcher watches the session directory with inotify so high-activity periods get captured faster than the cron interval
• Pre-compaction hook fires when OpenClaw is about to compact context, ensuring nothing is lost

Quick Start

1. Install the skill

clawdhub install total-recall

2. Set your API key

Add to your .env or OpenClaw config:

OPENROUTER_API_KEY=sk-or-v1-xxxxx

3. Run the setup script

bash skills/total-recall/scripts/setup.sh

This will:

• Create the memory directory structure (memory/, logs/, backups)
• On Linux with inotify + systemd: install the reactive watcher service
• Print cron job and agent configuration instructions for you to add manually

4. Configure your agent to load observations

Add to your agent's workspace context (e.g., MEMORY.md or system prompt):

At session startup, read `memory/observations.md` for cross-session context.

Or use OpenClaw's memoryFlush.systemPrompt to inject a startup instruction.

Platform Support

All core scripts use portable bash. stat, date, and md5 commands are handled cross-platform via _compat.sh.

Configuration

All scripts read from environment variables with sensible defaults:

LLM Provider Configuration

Total Recall uses any OpenAI-compatible chat completion API. Switch providers by setting environment variables:

Provider examples

# OpenRouter (default)
export OPENROUTER_API_KEY="your-key"

# Ollama (local)
export LLM_BASE_URL="http://localhost:11434/v1"
export LLM_API_KEY="ollama"
export LLM_MODEL="llama3.1:8b"

# Groq
export LLM_BASE_URL="https://api.groq.com/openai/v1"
export LLM_API_KEY="your-groq-key"
export LLM_MODEL="llama-3.3-70b-versatile"

Files Created

memory/
  observations.md          # The main observation log (loaded at startup)
  observation-backups/     # Reflector backups (last 10 kept)
  .observer-last-run       # Timestamp of last observer run
  .observer-last-hash      # Dedup hash of last processed messages
logs/
  observer.log
  reflector.log
  session-recovery.log
  observer-watcher.log

Cron Jobs

The setup script creates these OpenClaw cron jobs:

Reactive Watcher (Linux only)

The reactive watcher uses inotifywait to detect session activity and trigger the observer faster than cron alone. Requires Linux with inotify-tools installed.

# Install inotify-tools (Debian/Ubuntu)
sudo apt install inotify-tools

# Check watcher status
systemctl --user status total-recall-watcher

# View logs
journalctl --user -u total-recall-watcher -f

Cost

Using the default free models via OpenRouter:

• ~$0.00/month (using free-tier models) for typical usage (observer + reflector)
• ~15-30 cron runs/day, each processing a few hundred tokens
• All default models are free tier models

Model Notes

The scripts include defensive handling for both .content and .reasoning fields in API responses. While most models return content in the standard .content field, the fallback to .reasoning ensures compatibility with models that may use different response formats.

How It Works (Technical)

Observer

1. Finds recently modified session JSONL files
2. Filters out subagent/cron sessions
3. Extracts user + assistant messages from the lookback window
4. Deduplicates using MD5 hash comparison
5. Sends to LLM with the observer prompt (priority-based compression)
6. Appends result to observations.md
7. If observations exceed the word threshold, triggers reflector

Reflector

1. Backs up current observations
2. Sends entire log to LLM with consolidation instructions
3. Validates output is shorter than input (sanity check)
4. Replaces observations with consolidated version
5. Cleans old backups (keeps last 10)

Session Recovery

1. Runs at every /new or /reset
2. Hashes recent lines of the last session file
3. Compares against stored hash from last observer run
4. If mismatch: runs observer in recovery mode (4-hour lookback)
5. Fallback: raw message extraction if observer fails

Reactive Watcher

1. Uses inotifywait to monitor session directory
2. Counts JSONL writes to main session files only
3. After 40+ lines: triggers observer (with 5-min cooldown)
4. Resets counter when cron/external observer runs are detected

Customizing the Prompts

The observer and reflector system prompts are in prompts/:

• prompts/observer-system.txt — controls how conversations are compressed
• prompts/reflector-system.txt — controls how observations are consolidated

Edit these to match your agent's personality and priorities.

Dream Cycle

The Dream Cycle is an optional nightly agent that runs after hours to consolidate observations.md. It archives stale items and adds semantic hooks so nothing useful is actually lost. Context stays lean; everything remains findable.

What It Does

• Classifies every observation by impact (critical / high / medium / low / minimal) and age
• Archives items that have passed their relevance threshold
• Adds a semantic hook for each archived item (specific keywords + archive reference)
• Validates the result and rolls back automatically if something goes wrong