No manual tool calls. No "remember this" prompts. No extra setup.
It automatically: • Searches and injects relevant memories before every prompt • Logs every agent action and tool result (file reads, script executions, etc.) • Self-heals every 5 minutes • Keeps full Git-based audit trail • Resolves memory conflicts autonomously
Currently 100% stable and working out-of-the-box with Gemini CLI. Support for Claude Desktop and Cursor coming very soon.
Project: https://github.com/sl4m3/ledgermind
Would love feedback from people building long-running autonomous agents.