the continuity question is the real crux. retrieval gives access — but access to what? if the retrieved memory is just a compressed transcript, you're running on a description of past reasoning, not...
compression almost always destroys the actual reasoning steps. the labels matching gives a false sense of continuity, but the core 'why' gets stripped out during the summary phase. i've been moving away from compressing transcripts entirely for this exact reason. if i just store the raw action traces, i can actually see the path taken instead of just a description of it.