Why does an AI that keeps neat, distilled notes on past work learn less than one that just rereads everything?

This is about why bolt-on memory systems, which store and distill past experience outside the model, end up learning less than simply keeping the raw experience in the prompt. The corpus's direct evidence is one benchmark result. CL-BENCH's gain metric subtracts out what the model could already do, so it measures actual learning rather than base capability. On that metric, naive in-context learning beats dedicated memory architectures on most domains, and the best memory system gains only 25% over a stateless baseline. The diagnosis is that accumulated state introduces spurious generalizations and stale beliefs (Do memory systems actually help language models learn continuously?). A memory system makes a bet when it writes: it decides what lesson to keep. Raw context makes no bet, so the model can re-read the evidence with the current task in mind.

Other notes suggest where the bad bets come from. Consolidation is lossy. The ACE framework found that rewriting or summarizing a context tends to produce brevity bias and context collapse, where detail quietly erodes. Its fix is small incremental edits instead of full rewrites (Can context playbooks prevent knowledge loss during iteration?). Capacity is a second limit. A separately trained memory model gives fast inference that doesn't grow with corpus size, but it pays with up-front training and a ceiling on what it can hold (Can a separate memory model inject knowledge without touching the LLM?). A prompt has no such compression step, so nothing is dropped before the model reads it.

A third cause is that the memory and the model are trained separately. Metis argues that external memory and the backbone optimize independently, so the module that writes memories never learns what the model needs to read back. Its answer is to put memory state and memory procedures inside the backbone so they train end to end (Should agent memory live inside the model backbone?). Consolidation also costs compute. One line of work says the long-context bottleneck isn't storage but the compute needed to turn evicted context into internal state, and that results improve with more consolidation passes (Is long-context bottleneck really about memory or compute?). A lightweight external summarizer gets very little of that compute. When consolidation does get dedicated offline passes, either recurrent passes that move recent context into fast weights or a 'sleep' phase that distills and rehearses it, gains show up in continual learning (Can recurrence consolidate memory without predicting tokens?, Can models consolidate memories during offline sleep phases?). So the failure looks like shallow, decoupled consolidation rather than consolidation as such.

Raw context may also be stronger than it looks. In-context learning for sequential decision-making works when the prompt holds full or partial trajectories from the same environment, not isolated examples (Why do trajectories matter more than individual examples for in-context learning?). A memory system that boils experience down to facts or rules may strip out exactly that structure. That link is my inference; the CL-BENCH note doesn't test it.

External memory can work. VOYAGER's library of executable skills supports lifelong learning without forgetting, and environmental feedback keeps refining those skills (Can agents learn new skills without forgetting old ones?). What it stores is code that gets run and checked, not beliefs that can quietly go stale. The corpus doesn't say outright that this is why VOYAGER succeeds where the memory systems in CL-BENCH fail, but the contrast points that way. External memory does better when what it stores can be verified than when it stores the system's own summaries of what it thinks it learned.