Give an AI agent its own distilled lessons from past attempts, and it mostly ignores them, even with nothing else to go on.

This explores why agents treat summarized experience as optional, even when it's the only evidence they have. In a test across 10 models and 9 environments, corrupting an agent's raw experience (the actual past trajectories) changed its behavior a lot. Corrupting the condensed version, the lessons it had written for itself, barely changed anything. The agents were using the transcripts and ignoring the summaries of them (Why do LLM agents ignore condensed experience summaries?).

The same work gives three reasons, and they stack. First, summaries lose the details an agent needs to act, such as the exact command that worked or the exact error that appeared, so a summary often can't change the next step. Second, models favor whatever is in front of them right now over anything retrieved from memory. Third, the pretrained model already has strong priors, so a note saying "try X" competes with what the model thinks it knows. The corpus doesn't say what happens when nothing at all competes with the summary. But the third cause suggests the model's own priors are always in the room, so "only evidence available" never means uncontested.

This isn't inevitable, and the form of the memory seems to matter. AgentFly improves entirely through memory, with no weight updates, and reached 87.88% on GAIA. It's built around stored cases, closer to raw experience than digested lessons, with separate subtask and tool memories (Can agents learn continuously from experience without updating weights?). DeepAgent also compresses history, but into structured episodic, working, and tool schemas, and it lets the agent decide when to fold. The authors credit that structure and autonomy for avoiding the degradation that badly designed consolidation causes (Can agents compress their own memory without losing critical details?). So the problem looks like lossy, unstructured summaries more than compression as such.

It also fits a wider pattern of models having information and not acting on it. Models often accept false premises in a question even when direct questions show they know the correct fact (Why do language models accept false assumptions they know are wrong?). Yet small models can be trained to ground answers in a passage, quote it literally, and abstain when it's missing. That suggests deferring to provided context is something training builds, not something scale supplies (Can small models learn to ground answers in context?). Nobody seems to train agents to defer to their own summaries.

The practical lesson from the skills and harness work is to stop expecting the model to reread and obey a note. Skills help mostly as procedural anchors that stabilize action, not as facts injected into the prompt, and they fail when retrieved out of context or followed too rigidly (Do skills teach procedures or inject missing facts?). More broadly, reliability comes from putting memory, skills, and protocols into the surrounding system so the model doesn't have to remember to use them (Where does agent reliability actually come from?). A summary only counts as experience if changing it changes what the agent does. That is the test the first study ran, and most summaries fail it.