AI agents keep a summary of what they've tried before — but do they actually use it, even with nothing else to go on?
Do agents actually use condensed experience when it is their only available evidence?
This explores whether an agent's behavior changes when the only experience it has is a summary of past attempts, with no raw record and nothing else to lean on. The corpus answers the usual case clearly but has no direct test of this stripped-down case.
In the normal setup, agents mostly don't use condensed experience. Across 10 models and 9 environments, perturbing the raw record of past attempts changed agent behavior a lot, while rewriting the condensed summaries barely moved it Why do LLM agents ignore condensed experience summaries?. The paper gives three reasons. Summaries drop the details that matter, models favor what is in front of them over what was retrieved, and pretrained knowledge makes outside experience feel unnecessary.
The corpus doesn't say what happens when the summary is the only evidence. Two of those three causes are about competition from immediate context and from pretraining, so a summary with no competitor might get used more. That is my inference, not a tested result. The third cause, lost detail, doesn't go away when the competition does, so a summary that gets read can still leave the agent with too little to act on.
There is also evidence that condensing is the risky step. When an LLM keeps rewriting its own memory as experience piles up, the memory eventually does worse than simply keeping the raw episodes. In one test, GPT-5.4 failed 54% of problems it had previously solved after consolidation Does agent memory degrade when continuously consolidated?. One of the three mechanisms is applicability stripping: the lesson survives the summary but the conditions for using it don't. An agent that follows such a summary faithfully would be following it wrongly.
Memory that keeps case-level structure seems to fare better. AgentFly stores cases, subtasks and tool use as memory and improves without touching model weights, reaching 87.88% on GAIA validation Can agents learn continuously from experience without updating weights?. Condensed material from outside the agent's own history can help too. Skills distilled from tutorial videos and code repositories lifted output quality by 11.9 points Can tutorial videos teach software agents reusable skills?. What such skills do matters, though. Most of their benefit comes from acting as procedural anchors that stabilize what the agent does (65.7% of cases), not from supplying facts (4.5%), and they fail when retrieved wrongly or followed too rigidly Do skills teach procedures or inject missing facts?.
Put together, the pattern is not that condensed information is always ignored. Agents seem to ignore self-written prose lessons and to respond to procedure-shaped guidance and to case records that keep their detail. The way to check whether a summary is actually used is the first paper's method of changing its content and watching whether behavior changes, since its presence in the prompt shows nothing about use.