Agent Memory

Persistent agent memory stops repeated AI mistakes by recalling prior failure context and enforcing auto-generated prevention rules before the next tool action executes. This ensures past feedback actively blocks recurring unsafe commands and bad tool usage.

You capture structured feedback for MCP-compatible agents by logging thumbs-up or thumbs-down signals with specific context and required changes. This records actionable lessons for improvement tracking across runs.

Local-first agent memory stores and derives prevention rules from repeated down signals directly on your machine. It operates entirely without an API key, ensuring feedback capture and recall remain private and self-contained.

To start an MCP server for agent memory, add the ThumbGate package to your MCP setup and run the command `npx thumbgate serve`. This enables the recall and capture_feedback tools for your workflows.

Yes, you can use agent memory for regulated or high-liability workflows. It enforces structured success and failure logging, converts repeated mistakes into prevention rules, and aids incident recovery by recalling past failures.

Local recall prevents bad tool usage by relying on previously captured feedback and derived rules, but it requires consistent thumbs-down logging to build effective prevention patterns. It does not inherently detect novel errors not yet captured in local memory.