The Problem with Context Based Memory
Most LLM-powered systems hit a token limit and do the dumbest possible thing: they chop from the front. Take the first N characters, summarize it, discard the rest. Sometimes the logic is slightly smarter: random sample, drop from the middle, oldest-first. But the result is the same. You might silently lose the fact that the player killed the town’s mayor three sessions ago, while keeping five lines of small talk about the weather.<br> More