> Paper: *Useful Memories Become Faulty When Continuously Updated by LLMs* > Authors: Dylan Zhang et al. (UIUC, Tsinghua, UChicago, UWashington) > arXiv: https://arxiv.org/abs/2605.12978 | Research date: 2026-05-25
1. A Counterintuitive Experiment
GPT-5.4, with no memory, solved 19 ARC-AGI tasks with 100% accuracy.
The same tasks, the same model — but equipped with a system that "automatically summarizes experience into memory." After each task, the model compresses its solution process into a reusable "lesson" and stores it in a memory bank. After 10 rounds of updates, accuracy fell to 52.6%.
The model didn't get worse. The tasks didn't get harder. The memory itself was decaying.
This is the core finding of a recent paper from UIUC and collaborators. It undermines a widely accepted engineering assumption: automatic post-task summarization cannot be treated as a safe default.
2. Core Finding: Memory Utility Rises, Then Falls
The paper studies the common "consolidated memory" paradigm in agent systems — LLMs rewriting past trajectories into textual memories that are continuously updated, in pursuit of self-improvement without parameter updates.
2.1 A Non-Monotonic Curve
On ScienceWorld, Agent Workflow Memory (AWM) shows a typical curve over update rounds:
- Early: memory is empty; performance near the memory-free baseline
- Middle: memory accumulates; performance climbs to a peak
- Late: continued updating degrades performance, sometimes below the memory-free baseline
- Task pool: 6 known ARC-AGI problem types
- Inputs: ground-truth solutions available, ensuring every trajectory is "useful"
- Update modes: Stream (update per task) vs. Static (one-shot over the whole pool)
- Evaluation: same model on the same tasks before and after updates
- Episodic memory: rapid storage of individual experiences with details and context
- Semantic memory / schemas: slow extraction of abstract patterns supporting generalization
- Episodic Buffer: raw trajectories, retained individually
- Abstract Store: compressed lessons
- Auto mode: the agent decides when to retain or abstract
- Force mode: abstraction forced every round; raw trajectories may not be kept
- Abstract Only: never exceeds the memory-free baseline
- Episodic Only: recovers nearly all of Auto's gains
- Auto (both): only slightly above Episodic Only
- Prioritize retaining raw trajectories
- Trigger consolidation explicitly, not automatically
- Attach applicability conditions to abstractions; never strip context
- Support version rollback so faulty abstractions can be "forgotten"
- [ ] Implement an episodic buffer in OpenClaw-style MEMORY.md update flows
- [ ] Test abstraction-quality differences across consolidators (GPT-5.4 vs. Claude vs. Qwen) on the same trajectory pool
- [ ] Check whether parametric memory (vector DB embeddings) obeys the same dynamics
- [ ] Examine whether multimodal agents show the same degradation patterns
- [ ] Explore whether fine-tuned consolidators can fix current failure modes
WebShop is more extreme: AWM drops from 0.64 with 8 examples to 0.20 with 128, while the memory-free baseline sits at 0.20. Scaling memory eventually erases all of its own gains.
2.2 Even Strong Memories Degrade
Even when initial memories are generated by the strongest model (GPT-5.4) and then updated by GPT-5-nano, memory utility still decays. Cross-validation with Qwen3.5-{27B, 9B, 4B} solvers shows the same decline.
This means: memory quality depends not on the initial consolidator's capability, but on the update mechanism itself.
3. ARC-AGI Stream: The Cleanest Failure Case
The paper designed a controlled environment, ARC-AGI Stream, to eliminate confounds.
3.1 Setup
3.2 Key Numbers
| Condition | GPT-5.4 accuracy | |-----------|------------------| | No memory | 100% (19/19) | | Static-All (one-shot consolidation) | 94.7% | | Stream R10 (per-task updates, 10 rounds) | 52.6% | | Stream R50 (50 rounds) | Lower still |
Static consolidation held at 94.7% after both 10 and 50 rounds. Stream consolidation fell 47 points after just 10 rounds.
The difference lies not in the content but in the update mode. The same trajectory pool, consolidated once vs. incrementally, produces radically different memory quality.
4. Three Failure Mechanisms
4.1 Misgrouping
Consolidation begins by deciding which experiences belong together. LLMs often group structurally different trajectories and abstract across them, producing "lessons" that mix features from incompatible tasks. Group-wise consolidation (Static-Group) significantly outperforms mixed consolidation (Static-All).
4.2 Overgeneralization
Even with correct grouping, abstraction strips out applicability conditions. A lesson valid in one context becomes a universal rule that interferes with neighboring tasks. In task-switching experiments, a +203-point gap between "Fresh" and "Cumulative" consolidation groups came mainly from accumulated over-generalized and garbage memories.
4.3 Overfitting to Narrow Streams
When the input stream is narrow (e.g., near-duplicates of the same task), memory overfits to seen instances and fails to generalize on new in-task instances. Figure 11 shows memory performs fine on compressed problem sets but collapses out-of-distribution.
All three point to one conclusion: LLMs unreliably judge what to generalize, what to preserve, and what to discard. Each update overwrites the previous output, amplifying small errors.
5. Cognitive Science Foundations
The paper anchors its findings in dual-process memory theory.
5.1 Complementary Learning Systems
McClelland et al. (1995) describe two parallel memory systems in the brain:
They are complementary, not substitutive. But uncontrolled schema formation — non-selective integration, misgrouping — causes interference and loss of specificity.
5.2 The LLM Agent Deviation
Current agent memory systems have schema formation without episodic retention. Humans preserve raw episodes first; consolidation happens slowly, offline, while the original episodes remain retrievable. Consolidated-memory agents skip episodic retention and force immediate abstraction — akin to forcing the brain to complete all integration one second after an event and delete the original experience. The paper cites Bartlett (1932): humans retelling stories unconsciously fill gaps with their own schemas, drifting from the original narrative. LLM consolidation is the same mechanism — faster, larger-scale, and with no original text to check against.
5.3 Missing Metacognitive Control
Human consolidation involves metacognitive judgment: Is this worth remembering? At what granularity? How to handle conflicts? LLM agents set the consolidator and solver to the same model, asking it to monitor its own abstraction quality. Related research shows chain-of-thought explanations diverge from underlying computation (Turpin et al., 2023); a consolidator's self-assessment is similarly unreliable.
6. Engineering Judgment: Auto vs. Force
6.1 Design
The agent maintains two stores:
Three actions per step: Retain, Delete, Consolidate.
6.2 Results
| Condition | GPT-5.4 cumulative success rate, 400 steps | |-----------|--------------------------------------------| | No memory | 22.0% | | Force | 23.8% | | Auto + Episodic=50 | 35.5% | | Auto + Episodic=100 | 37.8% |
Ablations:
Raw trajectories carry all the effective information. Abstract memory adds no value on its own and only helps marginally in the right configuration.
6.3 Internal Dynamics of Auto Mode
The episodic store saturates: the agent converges to keeping a few representative trajectories per problem type across all 6 types. Correct grouping ability exists — it just requires autonomy and time.
Larger buffers mean less compression:
| Episodic capacity | Compress ratio | Keep ratio | |-------------------|----------------|------------| | 50 | 32.5% | 58.5% | | 100 | 20.8% (halved) | 60.8% (slightly up) |
This matches the "fast episodic + slow schema" dual-process theory.
7. Engineering Implications
7.1 The Default Is Wrong
Mainstream frameworks — CLIN, AWM, Dynamic Cheatsheet, ACE — all force memory rewrites after every interaction, assuming "more experience = better memory." That assumption is wrong. The correct default: raw trajectories are first-class evidence; abstraction is optional and must be explicitly gated.
7.2 Redesign Principles
1. Preserve raw episodes — don't treat all trajectories as raw material for compression 2. Selective consolidation — trigger abstraction only on clear signals 3. Recoverability — abstractions must trace back to source trajectories
7.3 Mapping to Existing Frameworks
| Framework | Memory strategy | Suggested adjustment | |-----------|-----------------|----------------------| | CLIN | Continuously updated textual memory | Add an episodic buffer; gate consolidation | | AWM | Automatically accumulating workflow memory | Distinguish raw trajectories from abstract lessons | | Dynamic Cheatsheet | Dynamic cheatsheet updates | Keep version history; allow rollback | | ACE | Continuous abstraction into reusable skills | Delay abstraction; keep raw evidence first |
The author also notes direct relevance to OpenClaw-style systems: MEMORY.md updates that overwrite old content after each session follow exactly the criticized paradigm. Suggested fixes: preserve retrieval of raw conversation records, keep historical versions instead of full overwrites, retain full decision context rather than only conclusions, and periodically audit memory accuracy rather than trusting automatic summaries.
8. Limitations
1. Benchmarks: validated only on text-based agent benchmarks (ALFWorld, ScienceWorld, WebShop, AppWorld, Mind2Web) and ARC-AGI Stream; embodied, multimodal, tool-heavy production settings unknown. 2. Memory form: only natural-language abstraction studied; parametric memory (weight updates, distillation) and structured non-text representations are out of scope. 3. Models: both consolidator and solver are current LLMs; stronger or fine-tuned consolidators may change conclusions. 4. Statistics: limited repetitions per task due to API costs; no formal error bars. Conclusions rest on cross-model, cross-benchmark, cross-framework consistency.
9. Conclusion
The paper's contribution is not denying the value of memory, but denying the default of mindless automatic updating. LLMs can distill useful patterns from experience, but the process requires metacognitive control that current LLMs lack.
Engineering judgment: automatic post-task summarization must not be a safe default. A more conservative design:
As the paper puts it: "Until agents can control when and how to consolidate experience, continuously updated textual memory should be treated not as a reliable engine of self-improvement, but as a fragile mechanism that can make more experience produce worse memory."
More experience, worse memory. Not a paradox — a consequence of the current architecture.
10. Open Questions
References
1. Paper: https://arxiv.org/abs/2605.12978 2. Frameworks: CLIN (Majumder et al., 2023), AWM (Wang et al., 2024), Dynamic Cheatsheet (Suzgun et al., 2026), ACE (Zhang et al., 2025) 3. Cognitive theory: McClelland et al. (1995) complementary learning systems; Tse et al. (2007) schema formation; Bartlett (1932) memory reconstruction 4. Chain-of-thought reliability: Turpin et al. (2023)