Comment by amit2403

The thing I did not expect, running a persistent multi agent system for about eight months, is that the summarizing is rarely what kills recall. What kills it is that everything you keep has equal weight forever. I had relationship scores between agents that only ever accumulated, and within a few weeks every pair was pinned at the maximum and the number stopped carrying any information at all. Adding decay and diminishing returns near the ceiling fixed more than any retrieval change did. So I would be curious whether never summarizing eventually produces the same flattening for you, just later and with more tokens.