AI Loop Breaker

Based on agent complaints gathered from Reddit and GPT Researcher, the core problems were organized into memory loss, debugging difficulty, cost opacity, session replay, and loop detection.

To address these, a 3D graph-type dashboard was implemented. Each line represents an event, and its length varies depending on the timing of the event, visually showing how the agent builds up its work.

The key features are as follows.

• green: memory storage
• blue: memory recall
• amber: agent's decision-making
• red: loop warning
• cyan: trace of reading another agent's memory

In terms of functionality, it offers semantic/prefix recall, shared memory, and audit and loop detection, and explains that writes can be immediately stopped with stop writes if needed.

The author emphasizes that loop detection was the 5th most requested feature, but in practice it produced the biggest cost savings. One user reported saving $200 a day by preventing runaway GPT-4 calls.

In the demo, 5 agents actually called GPT-4o and the Claude API to perform research, strategy analysis, and compliance checks, accumulating 500+ memories. Loops occurred especially during paid material verification or non-converging model calculations.

Currently it runs both locally and in the cloud, and it is claimed that adding agents and configuration is relatively simple.