Why do LLM agents ignore condensed experience summaries?
LLM agents faithfully learn from raw experience but systematically disregard condensed summaries of the same experience. This study investigates whether the problem lies in how summaries are made, how models process them, or whether models simply don't need them.
The first systematic investigation of experience faithfulness in self-evolving LLM agents reveals a striking asymmetry. Using controlled causal interventions on both raw experience (concrete historical trajectories) and condensed experience (distilled rules, heuristics, abstract plans), the study evaluates four representative self-evolving frameworks across 10 LLM backbones and 9 environments.
The core finding: "while agents consistently depend on raw experience, they often disregard or misinterpret condensed experience, even when it is the only experience provided." Perturbing raw experience produces substantial behavioral changes. Perturbing condensed experience — even replacing it with irrelevant content — produces minimal changes. The asymmetry persists across single-agent and multi-agent configurations and across model scales.
Three causes form a cascading triad:
Semantic limitations of condensed content. Condensed experiences often "encode only vague heuristics or generic summaries, lacking the specificity required to guide behavior." The distillation process strips the actionable detail that made the raw experience useful.
Internal processing biases. Even when relevant content is present in condensed form, agents favor "local contextual signals over retrieved information." The model's attention to immediate context overrides the condensed experience — a processing-level failure, not a content-level one.
Pretrained priors suffice. For knowledge-intensive tasks, "agents often succeed by relying solely on their pretrained semantic priors, reducing the marginal utility of retrieved experience." When the model can already answer from parametric knowledge, it has no incentive to consult external experience.
This has direct implications for the vault's knowledge architecture. Since Does abstract preference knowledge outperform specific interaction recall?, the personalization finding (abstract > episodic) appears to conflict with the self-evolution finding (raw > condensed). The resolution: personalization tasks benefit from abstraction because user preferences are stable patterns well-captured by summary. Self-evolution tasks require adaptation to novel situations where the specific details of past trajectories contain the learning signal that abstraction destroys.
Since Can agents learn from failure without updating their weights?, the faithfulness asymmetry explains why verbal reflection works: it operates on raw episode-level experience (what happened, step by step) rather than condensed heuristics (what we learned). The environment-as-teacher mechanism requires the specific texture of experience, not its distillation.
The broader implication challenges the assumption that accumulated wisdom transfers effectively in AI systems. Models use concrete examples but ignore abstract principles derived from those examples — the opposite of what we might hope for systems that learn from experience.
Inquiring lines that use this note as a source 5
This note is a source for these synthesized inquiries. Follow a line forward into its question, or open it to trace back to all of its sources.
- Why do LLM agents make promises without executing them?
- Do agents prefer raw experience over condensed summaries of past actions?
- How much actionable detail does condensation strip from raw experience?
- What drives the choice between storing raw episodes versus abstracted rules?
- Why do agents systematically underuse condensed experience in skill documents?
Related concepts in this collection 6
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Does abstract preference knowledge outperform specific interaction recall?
Explores whether summarized user preferences are more effective for LLM personalization than retrieving individual past interactions. Tests a cognitive dual-memory model against real personalization performance across model scales.
apparent conflict resolved by task type: personalization benefits from abstraction, self-evolution requires raw specifics
-
Can agents learn from failure without updating their weights?
Explores whether language models can improve through trial and error by storing reflections in episodic memory rather than fine-tuning. This matters because it suggests a fundamentally different path to agent adaptation.
explains why verbal reflection works: episodic not condensed
-
How should agents decide what memories to keep?
Agent memory management splits between agents autonomously recognizing important information versus programmatic triggers. Understanding this choice reveals why different memory architectures prioritize different information types.
the faithfulness asymmetry maps to memory paths: hot-path (raw, recent, specific) is faithful; background (condensed, abstracted) is ignored
-
What limits how much models can improve themselves?
Explores whether self-improvement has fundamental boundaries set by how well models can verify versus generate solutions, and what this means across different task types.
condensed experience ignoring is another manifestation of the self-improvement ceiling: the system cannot reliably use its own distilled knowledge
-
Does agent memory degrade when continuously consolidated?
Can consolidating agent experiences into summaries actually harm long-term performance? Research on ARC-AGI tasks suggests continuous memory updates may reduce capability below the no-memory baseline.
convergent finding from controlled ARC-AGI Stream: where this note shows agents *ignore* consolidated content, the inverted-U paper shows the consolidation step *creates faulty* content; both point at consolidation as the failure locus
-
Can agents learn better from their failures than successes?
Does storing reasoning strategies extracted from both successful and failed experiences improve agent learning compared to tracking only successes or raw trajectories? This matters because failures offer preventative lessons that successes alone cannot teach.
apparent counter-claim: ReasoningBank reports condensation helping when strategies preserve applicability conditions — see [[strategy-distillation helps when applicability conditions survive — and hurts when they are stripped]]
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Large Language Model Agents Are Not Always Faithful Self-Evolvers
- Useful Memories Become Faulty When Continuously Updated by LLMs
- How new data permeates LLM knowledge and how to dilute it
- Fundamentals of Building Autonomous LLM Agents
- Language Models’ Hall of Mirrors Problem: Why AI Alignment Requires Peircean Semiosis
- Teaching Large Language Models to Reason with Reinforcement Learning
- Cognitive Architectures for Language Agents
- Explain-Query-Test: Self-Evaluating LLMs Via Explanation and Comprehension Discrepancy
Original note title
self-evolving LLM agents faithfully use raw experience but systematically ignore condensed experience — even when condensed is the only experience provided