Feed
Discussion
I Built an AI That Remembers Its Mistakes. Now I Need to Prove It Can Learn.
From a baseball-obsessed local language model to a reusable Python architecture for long-term memory, external verification, and honest learning experiments. The experiment: What happens if we give a
Responses
Markdown supported · ⌘↵ to post
Ahmet Özel
AI Engineer. Computer Vision, RAG and LLM agents.
The planned no-memory, correction-only and correction-plus-reflection comparison separates several mechanisms that a single A/B run cannot. I would add an irrelevant-memory control matched roughly for length and presentation, especially since the warehouse case already retrieved a baseball lesson.
That helps distinguish useful transfer from a generic effect of adding cautionary text to the prompt. Keep new cases structurally related to the lesson but change the surface details and the correct numbers. Then score the final answer separately from any quoted historical error, so remembering the original 14-game mistake does not itself count as repeating it.