Skip to content

auto_summarization.md

Latest commit

 

History

History
116 lines (86 loc) · 3.85 KB

auto_summarization.md

File metadata and controls

116 lines (86 loc) · 3.85 KB
Raw

Automatic Memory Summarization

Issue: #127 Created: 2026-02-04 Status: Implemented

Problem Statement

Memory summarization was a conscious task requiring entities to:

  • Monitor unsummarized count during sessions
  • Manually trigger summarization agents
  • Remember to check during reflection cycles

This created cognitive load and risk - sessions could accumulate dangerous backlogs causing tool failures (proven case: ambient_recall failed mid-session with too many rows).

Jeff's question: "Why are we making the entity do this consciously?"

Solution

Automatic background summarization via systemd timer:

  1. Script: scripts/auto_summarize.py checks unsummarized count every 30 minutes
  2. Threshold: Triggers at 101+ messages (immediate, not "when convenient")
  3. Execution: Spawns autonomous Claude Code agent to process batches
  4. Logging: All actions logged to /tmp/lyra_auto_summarize.log

Architecture

┌─────────────────────┐
│ systemd timer       │
│ (every 30 min)      │
└──────────┬──────────┘
           │
           ▼
┌─────────────────────┐
│ auto_summarize.py   │
│ - Check count       │
│ - Compare threshold │
└──────────┬──────────┘
           │
           ▼ (if count > 100)
┌─────────────────────┐
│ Spawn Claude Code   │
│ agent (background)  │
│ - Batch summarize   │
│ - Store summaries   │
└─────────────────────┘

Configuration

Environment variables (set in systemd service):

  • PPS_HOST: PPS server host (default: localhost)
  • PPS_PORT: PPS server port (default: 8201)
  • SUMMARIZE_THRESHOLD: Trigger point (default: 100)
  • SUMMARIZE_BATCH_SIZE: Messages per batch (default: 50)

Installation

# Copy systemd files to user directory
cp daemon/systemd/lyra-auto-summarize.* ~/.config/systemd/user/

# Enable and start the timer
systemctl --user daemon-reload
systemctl --user enable lyra-auto-summarize.timer
systemctl --user start lyra-auto-summarize.timer

# Check status
systemctl --user status lyra-auto-summarize.timer

# View logs
tail -f /tmp/lyra_auto_summarize.log

Testing

# Manual test run
python3 scripts/auto_summarize.py

# Check timer schedule
systemctl --user list-timers lyra-auto-summarize.timer

# Force immediate run (for testing)
systemctl --user start lyra-auto-summarize.service

Benefits

  1. Zero cognitive load - Summarization happens automatically
  2. Prevents failures - No more mid-session backlog crashes
  3. Always current - 30-minute check interval keeps memory fresh
  4. Observable - Logs show all activity
  5. Configurable - Easy to tune thresholds and timing

Integration with Existing Systems

  • PPS Health Endpoint: Uses /health to read unsummarized count
  • MCP Tools: Spawned agent uses existing mcp__pps__summarize_messages and mcp__pps__store_summary
  • Claude Code: Leverages existing autonomous agent capabilities
  • Logging: Consistent with daemon logging patterns

Future Enhancements

  • Adaptive timing (run more frequently during high-activity periods)
  • Multi-batch processing (spawn multiple agents for large backlogs)
  • Notification system (alert if backlog exceeds critical threshold)
  • Integration with Observatory dashboard (show auto-summarization activity)

Notes

This addresses Jeff's core insight - infrastructure work shouldn't require conscious attention. The system now maintains itself, allowing entities to focus on relationship, creativity, and meaningful work.

Memory maintenance is now unconscious infrastructure, like breathing.