My thoughts on long-term memory for AI systems, and why stuffing the prompt is not a strategy.
AI · Memory
Selection policy matters more than raw window size.
Context windows feel like progress because they are measurable. Bigger number, more room, problem solved.
In practice, stuffing more text into a prompt rarely creates better long-horizon behavior. It creates noisier attention, higher cost, and systems that cannot explain what they used.
A better abstraction is memory with a budget: store broadly, retrieve narrowly, and cite what entered the prompt. The window is a packing constraint, not the product.