I built Bardacle to solve context loss in long-running AI agent sessions.
The problem: AI agents lose track of what they're doing during context compaction or session restarts. They forget the current goal, recent decisions, and why they were doing something.
Bardacle is a lightweight daemon that monitors agent sessions and maintains a "session state" summary using local LLMs. It tracks:
- Current goal
- Active tasks and their status
- Recent decisions
- Blockers
- Next steps
When context gets compacted, the agent can read this file to quickly re-orient.
Tech: Python, works with Ollama/LM Studio/Groq/OpenAI. Atomic writes, automatic backups, crash recovery. ~500 lines.
I built Bardacle to solve context loss in long-running AI agent sessions.
The problem: AI agents lose track of what they're doing during context compaction or session restarts. They forget the current goal, recent decisions, and why they were doing something.
Bardacle is a lightweight daemon that monitors agent sessions and maintains a "session state" summary using local LLMs. It tracks: - Current goal - Active tasks and their status - Recent decisions - Blockers - Next steps
When context gets compacted, the agent can read this file to quickly re-orient.
Tech: Python, works with Ollama/LM Studio/Groq/OpenAI. Atomic writes, automatic backups, crash recovery. ~500 lines.
Open source, MIT licensed. Feedback welcome!