AI agents rely heavily on their raw memories but barely notice when those memories are summarized into tidy lessons.
Why did agents ignore condensed experience in the memory rewrites?
This explores why LLM agents that learn from past experience barely change their behavior when that experience has been rewritten into summaries or distilled lessons, even though they lean heavily on the original raw records.
This explores why agents that learn from past experience barely change their behavior when that experience is rewritten into a condensed form. The most direct evidence is a study across 10 models and 9 environments. Changing the raw experience an agent was given shifted its behavior a lot, but changing the condensed version hardly mattered Why do LLM agents ignore condensed experience summaries?. The condensed memory was mostly decoration. The study names three causes: summaries lose critical details, models favor what's in front of them over what they retrieve, and pretrained knowledge makes the agent less dependent on any external experience.
The first cause may be the most important, because ignoring a summary can be the sensible choice. A separate study found that LLM-consolidated memory follows an inverted-U curve. It helps at first, then gets worse as experience piles up, and eventually does worse than just keeping the episodes. After consolidation, one frontier model failed 54% of the problems it had previously solved Does agent memory degrade when continuously consolidated?. The mechanisms were grouping unlike things together, stripping out the conditions under which a lesson applied, and overfitting to narrow streams of experience. Put the two findings together and an agent that discounts a lossy rewrite may be reacting to a real quality problem. That pairing is my reading, since neither study tests it directly.
The second cause is about how memory gets into the agent's thinking. Transcript replay and plain retrieval have no gating. Nothing decides what gets recalled versus what gets committed as permanent, so weak memory control, not missing knowledge, is what drives failures in long workflows Can agents fail from weak memory control rather than missing knowledge?. One alternative is to skip the fixed retrieve-then-reason pipeline and let reasoning walk through memory, pruning paths as evidence builds up. That approach gained up to 23% on reasoning tasks Can agents reconstruct memory on demand instead of retrieving it?. Memory the agent actively pulls on, rather than memory handed over as a block of text, may be harder to ignore.
Several notes point to fixes, though none tests whether they get agents to use condensed experience. Structure and autonomy help: agents that fold their own history into episodic, working, and tool schemas avoid the degradation of poorly designed consolidation Can agents compress their own memory without losing critical details?. So does matching the level of abstraction to the domain, with workflow-level summaries for routine tasks and causal rules or state-action records for others Does agent memory work better at one level of abstraction?. Letting links between memories form and get pruned based on whether they actually helped during execution also reached state-of-the-art on three benchmarks Should agent memory adapt dynamically based on execution feedback?.
To tell which of these causes is at work in a given system, you need to look inside the memory pipeline. Scoring only end-to-end task success hides whether extraction, storage, retrieval, or maintenance was the weak stage How should we actually evaluate agent memory systems?. An agent ignoring its summaries could be a problem with what the summary kept, or with how it was retrieved, or with the model's own habits, and that kind of evaluation is how you would separate them.
Sources 8 notes
Across 10 LLM models and 9 environments, perturbing raw experience changed agent behavior significantly, while altering condensed experience had minimal effect. Three causes drive this asymmetry: summaries lose critical details, models favor immediate context over retrieved information, and pretrained knowledge reduces reliance on external experience.
LLM-consolidated textual memory degrades as experience accumulates, eventually performing worse than episodic-only retention. GPT-5.4 failed 54% of previously-solved problems after consolidation, with three mechanisms identified: misgrouping, applicability stripping, and overfitting on narrow streams.
Agent performance degrades in long workflows because transcript replay and retrieval-based memory lack gating mechanisms. A bounded, schema-governed committed state that separates artifact recall from permanent memory write prevents error accumulation and constraint drift.
MRAgent achieves up to 23% gains on reasoning tasks by reconstructing memory through active graph traversal that prunes paths based on accumulated evidence, while reducing token and runtime cost compared to fixed-retrieval pipelines.
DeepAgent's autonomous memory folding consolidates interaction history into episodic, working, and tool memory schemas. This reduces token overhead while letting agents pause to reconsider strategies—the autonomy and structure together avoid degradation that plagues poorly designed consolidation.
Show all 8 sources
Workflow-level memory wins in routine-rich domains, causal-rule memory in environment-rich domains, and state-action memory in spatially-rich web tasks. The optimal abstraction depends on whether task variance comes from arguments, causal structure, or fine-grained UI state.
FluxMem demonstrates that adaptive memory topology—where links form, refine, and consolidate based on closed-loop execution feedback—consistently reaches state-of-the-art across three distinct benchmarks. Dynamic connectivity outperforms fixed retrieval by aligning abstraction and eliminating interference.
Decomposing memory into storage, extraction, retrieval, and maintenance stages exposes design trade-offs and failure modes that task-success metrics completely hide. Module-by-module evaluation across 12 systems shows which component actually failed rather than just whether the task succeeded.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Useful Memories Become Faulty When Continuously Updated by LLMs
- Are We Ready For An Agent-Native Memory System?
- Know It, Act on It: Investigating Memory Utilization in LLM Personalization
- Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory
- From Model Scaling to System Scaling: Scaling the Harness in Agentic AI
- GateMem: Benchmarking Memory Governance in Multi-Principal Shared-Memory Agents
- AI Agents Need Memory Control Over More Context
- Demystifying Agent Skills: Why They Work-Until They Don't