1 comments

  • sg17gweedo 5 hours ago ago

    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!