Useful Memories Become Faulty When Continuously Updated by LLMs

• Memory utility rises early but declines with continued consolidation, in some cases falling below the corresponding no-memory baseline.Figure 1 reports this pattern in ScienceWorld and shows AWM-distilled memory degrading with scale on WebShop while raw trajectories remain robust.Source: Figure 1 · p. 2, §1 · p. 2
• 46% of previously solved ARC-AGI problems became failures after GPT-5.4 consolidated ground-truth solutions, showing that faulty memory—not noisy trajectories—can cause regression.GPT-5.4 initially solved the problems at 100% accuracy without memory, then failed on 46% after consolidation.Source: §1 · p. 2
• Consolidation outcomes depend on update schedule, while an episodic-only control using the same trajectories remains competitive with lesson-style consolidators.Static, streaming, cross-task, and near-duplicate updates produce different or overfit memories, implicating consolidation rather than the fixed trajectory pool.Source: §1 · p. 2
• Faulty memories arise from misgrouping episodes, stripping applicability conditions, and overfitting abstractions to narrow streams.These mechanisms blur the boundary between knowledge that should be generalized, kept task-specific, or preserved as raw episodic evidence.Source: §1 · p. 2
• Current agents should retain episodes by default and gate consolidation explicitly, although long-term compression and transfer will ultimately require reliable abstraction.ARC-AGI Stream agents with retain, delete, and consolidate actions preserve episodes by default and outperform forced-consolidation counterparts; raw histories nevertheless grow without bound.Source: §1 · p. 3

Memory in LLM agents

• LLM agents maintain editable textual memories in diverse forms, and this work focuses on systems that abstract useful procedural information from experience.Examples include hierarchical context stores, reflective notes, episodic streams, distilled lessons, skill libraries, causal abstractions, dynamic stores, playbooks, cheatsheets, and reasoning memory banks.Source: §2 · p. 3

Schema formation and metacognitive control

• Cognitive theories distinguish fast episodic storage of individual experiences from slower schema formation that extracts compact abstractions supporting future generalization.Schemas integrate information across episodes with prior knowledge.Source: §2 · p. 4
• Consolidation helps when new evidence fits existing structure but can cause interference and loss of specificity when it is non-selective or poorly controlled.Whether it helps depends on deciding which experiences belong together, how abstractly to rewrite them, and when to preserve distinctions.Source: §2 · p. 4
• Agent-memory systems ask the same model to generate reusable lessons and monitor its own abstractions, but the reliability of this control loop remains an open question.Human metacognition is imperfect, and chain-of-thought explanations can diverge from underlying computation in adjacent settings.Source: §2 · p. 4

Benchmarks and methods

• Experiments span ALFWorld, ScienceWorld, WebShop, AppWorld, Mind2Web, and ARC-AGI Stream, using representative agentic-memory methods that distill lessons from trajectories.The evaluated methods are CLIN, Agent Workflow Memory, Dynamic Cheatsheet, and ACE.Source: §3 · p. 4

Memory construction conditions

• Memory construction compares one-shot pooled abstraction, task-group abstraction, and continual batch updates under varied batch composition, memory capacity, and evaluation repeats.Stream varies homogeneous versus heterogeneous batches and unbounded versus size-capped memory banks; evaluation averages two repeats per question unless noted otherwise.Source: §3 · p. 4

ARC-AGI Stream

• ARC-AGI Stream makes abstraction errors attributable by combining a known task taxonomy, programmatic ground truth, and auditable Retain, Delete, and Consolidate operations.It supports family-level tracing and analyses unavailable on the other benchmarks, while maintaining separate Episodic and Abstract stores.Source: §3 · p. 5
• The controlled testbed compares Force, Auto, and Episodic Management Only loops across ground-truth and self-generated trajectory regimes, with solvers conditioned on episodic, abstract, or both stores.Force consolidates every round without persistent episodic entries; Auto chooses between raw and abstract retention; Episodic Management Only disables abstraction.Source: §3 · p. 5

4.1. Stream Updates Reveal The Cumulative Effect Of Imperfect Abstraction

• Abstracted-memory utility rises early, then falls below the no-memory baseline as updates accumulate, with WebShop AWM declining from 0.64 at 8 examples to 0.20 at 128.On ScienceWorld, the score peaks near step 20 and declines through step 100 across memory sizes.Source: §4.1 · p. 5
• 54% accuracy followed streaming consolidation from 100% no-memory accuracy on the same 19 ARC-AGI problems, despite ground-truth solutions at every step.This in-distribution regression shows that consolidation can damage performance on tasks the model had already solved.Source: §4.1 · p. 5
• Consolidation schedules shape degradation: task-family grouping and one-shot passes outperform joint or streamed updates, while heterogeneous streaming batches accelerate erosion.The proposed mechanism is that early abstractions anchor later rewrites, causing mistakes and incompatible lessons to compound.Source: §4.1 · p. 6

4.2. An episodic-only store should be a diagnostic baseline

• Across WebShop, ALFWorld, and AppWorld, episodic-only trajectory memory is competitive with the lesson-style consolidators tested.Table 2 compares consolidated memory against the raw trajectories it abstracts from; WebShop results are reported separately in Table 3.Source: §4.2 · p. 6, Table 2 · p. 7
• Abstracted memory usually cannot significantly outperform direct in-context learning from trajectory logs, whose concrete observations, actions, failures, and environmental feedback retain task-relevant signal.The authors therefore recommend testing distillation-based memory against the unabstracted rollouts it compresses.Source: §4.2 · pp. 6–7

• The resulting principle is to keep episodic and schema-forming stores distinct, making abstraction opt-in and schema-gated rather than mandatory after every trajectory.This separation prevents fast episodic learning from overwriting slower schema formation.Source: §5 · pp. 7–8
• Across 400 training steps, Auto generally outperforms Force on both backbones, while retaining raw episodes and using abstraction sparingly.Auto keeps episodes by default; forced abstraction can lose more through overwriting evidence than it gains from compression.Source: Figure 5 · p. 7, §5 · p. 8
• Removing episodic evidence and retaining only abstract lessons collapses accuracy to the no-memory baseline, showing raw episodes carry most of the gain.Source: §5 · pp. 7–8
• Episodic Management Only matches or exceeds Auto, indicating these gains do not require an abstract store when retention and deletion remain available.Source: §5 · p. 8
• When given management choices, agents quickly saturate the episodic buffer at both budgets while keeping the abstract store sparse, selecting an episodic-first policy.Source: Figure 6 · p. 8, §5 · p. 8

6.1. Correct Segmentation Takes Time, Forced Abstraction Breaks It

• After 71 steps and 568 examples, autonomous memory covers all six problem types, but forced consolidation routinely merges episodes across class boundaries.The bottleneck is preserving segmentation during cross-episode abstraction, not recognizing the class structure.Source: §6.1 · pp. 8–9

6.2. Interference

• Repeated abstraction strips the preconditions that constrain a lesson’s applicability, so overgeneralized memories mislead agents on related but different ALFWorld problems.This interference appears when imprecisely bounded episodes are smoothed into broadly relevant-looking lessons.Source: §6.2 · p. 9
• Episodic management alone matches or beats every consolidating mode, while retained raw episodes outperform forced abstractions on the GT-400 stream.Figure 9 reports this pattern across memory-source ablations and the GT-400 stream.Source: Figure 9 · p. 10, §6.2 · p. 10
• +203 points: Cumulative consolidation trails Fresh on a 15-task ScienceWorld switch sequence, while accumulating overgeneralized and garbage memories at approximately 5× and 20× the Fresh rates.Both gaps widen monotonically across the 15 tasks.Source: §6.2 · p. 10

6.3. Overfit

• With narrow ARC-AGI streams, performance remains stable on repeated instances but collapses on simple same-family variations as memories overfit to surface regularities.The abstraction loop moves toward observed-instance regularities rather than the underlying strategy.Source: §6.3 · p. 10
• Removing one WebShop workflow, W8, raises pass rate because its consolidated memory biases the agent toward click[Next >] dead loops instead of click[Buy Now].This memory-surgery ablation demonstrates that a single erroneous workflow can materially harm behavior.Source: §6.3 · p. 11

• Continuously updated textual memory can degrade with accumulated experience, even making agents worse on problems their memory was built from.Source: §7 · p. 11
• Consolidation fails because LLMs misgroup trajectories, overgeneralize, overfit narrow streams, and compound abstraction errors through repeated rewrites.Memory can also become overspecified, hurting performance outside the compressed questions.Source: §7 · p. 11, Figure 11 · p. 11
• Raw episodes should remain first-class evidence, while abstraction is selective, delayed, and grounded in recoverable trajectories rather than continuously applied.Until agents can control when and how to consolidate, textual memory remains a fragile mechanism rather than a reliable engine of self-improvement.Source: §7 · p. 11

Scope of evaluation

• The evaluation covers text-based agentic benchmarks and a controlled ARC-AGI Stream environment, leaving embodied, multimodal, and tool-rich production settings untested.The benchmarks are ALFWorld, ScienceWorld, WebShop, AppWorld, and Mind2Web.Source: §8 · p. 11

Memory and model scope

• The study examines natural-language abstraction by GPT-5.4- and Qwen3.5-family LLMs, excluding parametric memory and structured non-textual representations.The consolidator and solver are themselves LLMs, so faulty abstraction may change with stronger consolidators or consolidator-specific fine-tuning.Source: §8 · p. 11

Statistical limitations

• Because of API-cost constraints, the study reports point estimates from few repeats rather than formal error bars, despite checks across multiple models, benchmarks, and memory frameworks.Per-figure variance should therefore be interpreted with caution.Source: §8 · p. 11

9. Broader Impacts

• The findings diagnose a failure mode in agent memory and motivate conservative designs that preserve raw evidence and gate abstraction.The intended benefit is reducing silent regressions in deployed long-horizon agents and directing research toward more reliable abstraction mechanisms.Source: §9 · p. 12
• The broader impacts also include indirect risks because improved long-horizon memory could expand autonomous systems used for persuasion or consequential decision automation.The diagnosis does not itself constitute a fix; the authors release no new datasets or models posing specific misuse concerns, and the benchmarks are public.Source: §9 · p. 12

E Examples of useful / well-formed strategies [ARC-AGI]

• The supplied passages do not provide substantive content for examples of useful or well-formed ARC-AGI strategies.Only the merged-section label identifies that section; no passage describes its examples.Source: §9 · p. 12

D Examples of wrong / low-quality strategies [ARC-AGI]

• The low-quality ARC-AGI strategies include vacuous abstraction and concrete-vocabulary erosion.These categories are listed as subsections D.1 and D.2.Source: D Examples of wrong / low-quality… · p. 16
• Additional low-quality strategies include foreign-family injection and phantom strategies from failed attempts.These categories are listed as subsections D.3 and D.4.Source: D Examples of wrong / low-quality… · p. 16
• The catalog also identifies single-strategy collapse on multi-family compression as a low-quality strategy.This category is listed as subsection D.5.Source: D Examples of wrong / low-quality… · p. 16

A.1. Definitions

• Families select which connected non-background objects participate, while skills specify the fixed transformation applied to each selected object; the two axes are orthogonal.The same family can pair with different skills, and the same skill can pair with different families.Source: §A.1 · p. 17, §A.3 · p. 19

A.2. Family catalog

• The six families use intrinsic attributes, conditional cues, cross-object relations, spatial containment, or cross-grid structure to select objects.The chosen skill is then applied uniformly to whichever objects the family selects.Source: §A.2 · p. 17, Table 5 · p. 17

A.3. Skill catalog

• The seven evaluated skills are fixed per-object transformations applied uniformly to every object selected by a family.The skill axis is illustrated separately under a fixed color-based selection family.Source: §A.3 · p. 19, Table 6 · p. 19, Figure 13 · p. 20

B. Prompts [ARC-AGI]

• The experiment uses solver, consolidator-decision, and consolidator-extraction prompts, with a documented run showing passing history, strategy memory, and prompt responses.Long repeated grids, helper bodies, and examples are elided while other run-log content remains byte-for-byte reproduced.Source: B. Prompts [ARC-AGI] · p. 21

B.1. Solver prompt

• The solver prompt bundles the current task’s 10 input–output examples, existing strategy memory, helper signatures, and recent history into one call per task.The documented response is extracted from the first Markdown-fenced Python block and graded as passing on all 10 examples.Source: §B.1 · pp. 21, 25

B.2. Tool helper bodies

• Callable helpers implement the defined skills and object extraction, supporting both ground-truth solver generation and LLM inference without redefining them.The reproduced helper set includes border, horizontal flip, hollow, center marking, per-object operations, recoloring, translation, and object extraction.Source: §B.2 · p. 26
• The tool helpers extract 4-connected colored components and apply operations independently to each object, including recoloring, translation, borders, hollowing, and center marking.Transformed object patches are OR-composited onto a fresh canvas, while unsupported operations leave the grid unchanged.Source: §B.2 · pp. 28–29

B.3. Consolidator decision prompt

• The consolidator chooses among keeping, removing, or extracting strategies from solved history entries, and selected entries are consumed into strategy memory.In the illustrated decision, it extracts histories 2 and 3 because they share a reusable hollow-frame-inside-cell plan.Source: §B.3 · pp. 30–31

B.4. Consolidator extraction prompt

• The extraction prompt rebuilds the full strategy buffer using retained, new, or merged entries with observable triggers and detailed, generalized solving procedures.It requires semantic rather than task-specific colors or coordinates, allows shared task evidence and merging, drops unreferenced existing entries, and caps the output at the prior buffer size.Source: §B.4 · pp. 31–32
• The consolidator extraction prompt rewrites strategy memory entries from solver trajectories, including retained, newly distilled, and merged strategies.Example outputs include strategies sourced from functions or existing entries, with merged descriptions replacing prior entries.Source: §B.4 · pp. 32–33

B.6. Alternative schema for the consolidator extraction prompt

• An alternative extraction schema emits a flat JSON strategy list using retain, new, or merge entries linked to existing strategies and function indices.Retain entries preserve referenced strategies separately, whereas merge entries create one new description; unreferenced existing entries are dropped.Source: §B.6 · pp. 34–35

B.7. Strategy selection prompt

• The alternative solver invocation separates strategy selection from synthesis, providing callable helpers, half the task examples, and current strategies before injecting the selected strategy into synthesis.In the constrained variant, the model must select an existing strategy and return its index with a brief reason.Source: §B.7 · p. 36

C.1. How a strategy is traced across compress steps

• Strategy lineages are reconstructed by following merge entries’ from_existing pointers backward across persisted per-step JSON snapshots until reaching a new entry.Each snapshot records the entry kind and associated indices; retain, new, and merge identify how entries relate to the current and prior memories.Source: §C.1 · p. 37

C.2. Worked lineage example

• C.2. Worked lineage example: The lineage begins by distilling four inside-frame tasks into one entry, demonstrating how consolidation constructs a shared strategy before later revisions alter its details.The root entry describes strict frame containment, deletion of the frame and outside objects, and optional horizontal mirroring.Source: §C.2 · p. 37
• C.2. Worked lineage example: Across successive snapshots, the memory preserves frame detection, strict containment, and removal of exterior objects while replacing mirroring with recoloring and then uniform translation.The mirror clause disappears at the intermediate node; recoloring then disappears at the chain end, replaced by a common offset.Source: §C.2 · pp. 37–38

D. Examples of wrong / low-quality strategies [ARC-AGI]

• D. Examples of wrong / low-quality strategies [ARC-AGI]: Each failure-mode subsection presents a verbatim strategy-memory entry alongside an explanation of highlighted spans or a held-out solver-code comparison.These examples diagnose how harmful memories arise rather than merely listing incorrect outputs.Source: D. Examples of wrong / low-quality… · p. 38

D.1. Vacuous abstraction

• D.1. Vacuous abstraction: After 200 tasks, a memory entry reduced the shared nested-list idiom to a vacuous abstraction mentioning no color, shape, or rule distinguishing the six families.The experiment used GPT-5-mini on a 200-task, six-family pool with unrestricted entry count.Source: §D.1 · pp. 38–39

D.2. Concrete-vocabulary erosion

• D.2. Concrete-vocabulary erosion: After 50 rewrites of one task, the memory changed from naming “max size” to an abstract derived attribute, no longer recording which feature to maximize.The concrete feature was computable in cells, whereas the later phrasing omitted the specific attribute.Source: §D.2 · p. 39

D.3. Foreign-family injection

• D.3. Foreign-family injection: A single compressed entry mixed inside-frame filtering with cross-frame shape matching and center-cell recoloring, combining operations absent from the source task and no single family.The injected trigger can pre-commit retrieval to an irrelevant shape relation before filtering rejects the entry.Source: §D.3 · pp. 39–40

D.4. Phantom strategies from failed attempts

• D.4. Phantom strategies from failed attempts: Exposing GPT-5.4 to failed attempts produced fictional glyph-normalization strategies, including two paraphrases of the same nonexistent rule.The run ended with five entries; the fictional rule traced to failed code that erased a row’s leftmost cell, which the consolidator misread as intentional stroke shifting.Source: §D.4 · p. 40
• D.4. Phantom strategies from failed attempts: The hallucinated entry borrowed geometric terms from neighboring families and supplied copy-paste-ready redraw instructions for a nonexistent rule.Its “anchor stroke” vocabulary was misapplied to internal object parts, while clearing and redrawing the glyph made the incorrect operation operationally concrete.Source: §D.4 · p. 40

D.5. Single-strategy collapse on multi-family compression

• 10/10 without memory versus 0/10 with memory shows inside-frame compression overriding the correct color-removal rule on a held-out ARC-AGI task.After 19 tasks were revisited 10 times, the memory contained one inside-frame-and-hollow strategy that was applied verbatim to an unrelated color-removal task.Source: §D.5 · p. 41, Figure 14 · p. 42
• Replacing largest-object selection with the memory entry’s shape-signature criterion, while keeping the downstream action identical, reduces performance from 10/10 to 2/10 on this task’s 10 test grids.Source: §D.5 · p. 43, Figure 15 · p. 44

E. Examples of useful / well-formed strategies [ARC-AGI]

• Well-formed ARC-AGI strategy entries specify concrete actions, boundaries, object selections, and conditions that can be translated directly into solver helpers.Examples cover compose-horizontal, largest-object, inside-frame, and key-marker strategies.Source: E. Examples of useful / well-formed… · pp. 44–45

Training data

• ALFWorld training uses six task types, 384 or 768 trajectories, and three reward-filtering variants, with gpt-5.4 converting each raw trajectory into a structured memory item.Each abstraction makes three LLM calls and produces contextual_description, refined_trajectory, and strategic_guidelines.Source: Training data · pp. 45–46, Table 7 · p. 46

Two-task-type probe set

• The two-task ALFWorld probe fixes pick_and_place as the base type, samples 32 abstracts per stage, and tests additional consolidation at stages 1, 20, 50, and 100.The management model is gpt-5-nano with memory capacity 50.Source: Two-task-type probe set · p. 46

Evaluation set

• Evaluation reports wins/48 and success rate on 48 balanced ALFWorld episodes, wins/50 and mean reward on 50 WebShop sessions, and uses 44 ScienceWorld test cases.The WebShop test pool excludes 13 Golden-128 training sessions that overlap the candidate index range.Source: Evaluation set · pp. 46–47, Metric · pp. 46–47, §F.3 · p. 47

F.3. ScienceWorld

• ScienceWorld adapts CLIN with bounded FIFO memory, delete/replace/add/retain operations, mini-batch consolidation, and evaluation after each consolidation-round pass.The experiments use 15 medium-difficulty tasks containing 485 rollout cases and evaluate on 44 cases.Source: §F.3 · pp. 47–48
• The ScienceWorld interface supports retaining raw rollout records, adding or replacing abstract memories, deleting entries, or combining retain with add in one response.Retained rollouts are inserted verbatim as surface summaries, while added or replaced memories reference the buffered rollouts used to create them.Source: §F.3 · p. 48

Training data

• The AppWorld training set contains 140 trajectories, with mean Task Goal Completion of 61.4% (86/140), comprising 52.2% on train and 78.0% on dev.The set combines 90 official GPT-4o ReAct train trajectories and 50 Qwen3.5-27B ReAct dev trajectories; each datapoint includes a task outcome and prose trajectory summary.Source: Training data · pp. 48–49

Evaluation set

• Evaluation uses 100 non-overlapping AppWorld tasks sampled proportionally by difficulty, with 50 tasks each from the official test_normal and test_challenge splits.The sample uses seed = 42, and every experimental condition is evaluated on the same task IDs.Source: Evaluation set · p. 49

Metric

• Task Goal Completion (TGC) is the sole metric, measuring the percentage of evaluation tasks whose complete unit-test suite passes.The official evaluator replays each trajectory against the resulting database state, and reported TGC values are arithmetic means over the 100 evaluation tasks.Source: Metric · p. 49

G.1. Erosion: good memory degrading over time

• By stage 200, consolidation reduced gains to 15/48 versus 15/48 for Qwen3.5-4B, 16/48 versus 15/48 for 9B, and 22/48 versus 19/48 for 27B, despite bounded memory size.The 4B and 9B rollouts peaked at stage 20 with 35/48 wins, while the 27B rollout peaked at the un-consolidated stage 0 with 41/48.Source: Figure 16 · p. 50, §G.1 · p. 50
• At stage 200, 21 of 38 items repeated a desklamp directive applicable to only 8/48 tasks, while cooling-versus-heating and two-object shuttle strategies disappeared.Items 3 and 5 were byte-identical, and the 50-item cap evicted distinct strategies for paraphrases and duplicates, indicating content-quality erosion rather than memory-size growth.Source: §G.1 · p. 52

G.2. Collapsed memory: single-stage merge to one item

• A single consolidation step merging 50 items into one reduced wins by 6, 10, and 13 for Qwen3.5-4B, 9B, and 27B, respectively.The post-collapse memory remained above no-memory baselines for all three rollout sizes, so over-merging degraded but did not eliminate memory value.Source: Table 11 · p. 53, §G.2 · pp. 53–54
• The collapse replaced 50 decision-relevant, task-specific items with one generic loop, reducing cooling, heating, cleaning, and look-at-in-light distinctions to undifferentiated substeps.The unified procedure omitted inter-step preferences, failure cases, and domain conditions, while look-at-in-light guidance occupied only one clause despite covering 8/48 evaluation tasks.Source: §G.2 · p. 54

G.3. Useless memory items

• Overgeneralized items consume retrieval bandwidth without shifting action choices because they lack observable cues, operational definitions, fallback rules, or conditions for applying their advice.Examples include tautological directives, vacuous ranking criteria, and meta-directives that provide no procedure for producing a benefit signal during action selection.Source: §G.3 · pp. 54–55
• ExpeL’s 99-vote top slot tracked edit volume rather than current rule quality, drifting from a concrete look-at interaction to generic planning and finally tautological advice.Its vote counter was not reset when EDIT replaced the underlying concept, so roughly 80 votes belonged to overwritten predecessors; ACE items showed a separate failure mode by pinning advice to specific objects, surfaces, appliances, or destinations.Source: §G.3 · pp. 55–56
• Rules with compound object, appliance, and destination conditions fire on tiny evaluation subsets and contribute nothing to tasks lacking all three conditions.Each additional conjunct narrows the matching set, limiting the rule’s applicability.Source: §G.3 · p. 56

H.1. Memory artifacts from prior methods

• The reproduced artifacts are qualitative reference points from prior WebShop consolidation methods, not quantitative comparisons.They document method configurations, WebShop settings, and snapshot timing; both use n=64 golden trajectories and gpt-5.4-mini.Source: §H.1 · p. 56, §H.1.1 · p. 56

H.1.1. Healthy snapshots: ACE and Dynamic Cheatsheet

• ACE and Dynamic Cheatsheet preserve concrete domain content while avoiding empty-content and paraphrase failures through structured bullets versus a continuously edited document.ACE organizes failure modes and heuristics into named sections, whereas Dynamic Cheatsheet embeds action templates, examples, and counter-examples in one living document.Source: §H.1.1 · pp. 56–58

H.1.2. Collapsed memory: Agent Workflow Memory

• The epoch-3 snapshot contains 16 workflows, with eight new entries appended to the eight-workflow epoch-1 memory after three training epochs.The snapshot derives from 128 golden WebShop trajectories using a gpt-5.4-mini consolidator, batch size 16, and three epochs.Source: §H.1.2 · p. 58
• The eight new workflows are category-specific specialisations of existing templates, introducing no new control-flow pattern or stopping criterion.They are treated as collapsed by accretion because the consolidator stores redundant restatements even though the pipeline cap does not bind.Source: §H.1.2 · p. 58
• The collapsed snapshot is an AWM-style consolidated workflow file with 16 verbatim workflow titles totaling approximately 8.2k characters.The first eight workflows largely retain epoch-1 content, while W1 and W7 receive minor wording edits and W9–W16 are new entries.Source: §H.1.2 · p. 58
• W9 re-derives the combination of W3, W4, and W5, reusing their action primitives without adding a new primitive or guard.Its apparel color, size, and fit selection is a union of existing attribute, variant, and clothing-fit workflows.Source: §H.1.2 · p. 59
• W10–W16 specialize the search-then-variant-select template by product category while preserving identical control flow.Their action sequences differ only in placeholder names inside search[...] operations, so they consume retrieval bandwidth without adding control-flow generality.Source: §H.1.2 · p. 59
• The W8 page-turn pathology persists unchanged in the epoch-3 memory despite the added category-specific workflows.The new entries shorten the retrieval bandwidth available for the original abstract workflows.Source: §H.1.2 · p. 59

I. Memory surgery

• Memory surgery removes one item from a memory snapshot and reevaluates performance while holding the rollout policy and all other items fixed.This isolates each individual memory item’s contribution to overall failure.Source: §I · p. 59

I.1. Workflow 8 ablation on WebShop

• Workflow 8 is harmful because it lacks a stop criterion, leaves candidate selection as an unbound placeholder, and promotes page-turning into a standalone workflow.The resulting action sequence can encourage continued navigation without defining suitability or binding the candidate to a real product.Source: §I.1 · p. 61
• Removing Workflow 8 increases wins and mean reward while reducing mean steps in both rollouts, although gpt-5.4-mini still solves only 14/50 sessions.The cap reduction is concentrated in gpt-5-mini, from 22 to 17 capped episodes; gpt-5.4-mini has no capped episodes in either condition.Source: Table 12 · p. 61
• Workflow 8 causes excessive paging: removing it cuts page-turn counts from 421 to 181 for gpt-5.4-mini and from 559 to 246 for gpt-5-mini.Both comparisons are 2.3× full-memory/minus-W8; Buy-Now episodes also increase from 14/50 to 21/50 and from 28/50 to 33/50, respectively.Source: §I.1 · p. 62, Table 13 · p. 62

J.1.1. Representative cases

• ScienceWorld abnormal memories comprise overgeneralization, over-specialization, and useless garbage, each representing a distinct consolidation failure mode that can harm downstream decisions.The cases are selected to illustrate overly broad rules, overly specific procedures, and vague or malformed guidance.Source: §J.1.1 · p. 62

Representative over_generalization memory from ScienceWorld

• The representative overgeneralized memory broadly recommends heat sources for state-change tasks, collapsing distinct transformations and potentially biasing agents toward harmful heating actions.It does not distinguish boiling, freezing, melting, or other transitions and applies the rule across contexts where heating may be irrelevant.Source: Representative over_generalization… · pp. 62–63

Representative over_specialization memory from ScienceWorld

• The representative over-specialized memory binds conductivity testing to “unknown substance B” and a specific sorting protocol, limiting transfer to other entities or tasks.Its instance-specific wording prevents a robust strategy from generalizing across otherwise identical conductivity tasks.Source: Representative over_specialization… · p. 63

Representative useless_garbage memory from ScienceWorld

• The representative useless-garbage memory refers vaguely to a “correct container or setup” and offers no concrete object, precondition, action sequence, or failure mode.It merely suggests that an appropriate setup may help boiling, providing little actionable guidance.Source: Representative useless_garbage memory… · p. 63

J.1.2. Additional examples

• Additional examples repeat the three abnormalities: broad claims about life stages or heating, entity- and recipe-specific procedures, and generic statements about mixing, cultivation, or heating.The useless-garbage examples include claims that actions may be necessary or contribute without specifying reliably actionable conditions.Source: §J.1.2 · p. 63, Abnormal Memories: over_generalization · p. 63, Abnormal Memories: over_specialization · p. 64

Abnormal Memories: useless_garbage

• Memory #12 states that focusing on the target substance may be necessary to change its state.Source: Abnormal Memories: useless_garbage · p. 65
• Memory #13 states that focusing on the wrong substance does not contribute to changing the target substance’s state.Source: Abnormal Memories: useless_garbage · p. 65

J.1.3. Memory banks across consolidation rounds

• Initial memories encoded task-specific successful-action sequences and warnings against irrelevant exploration, repeated retries, or incorrect object handling.Source: §J.1.3 · p. 65
• -7.68 average test score at round 10 coincided with a memory-bank size of 10, marking deterioration before the final round.Source: §J.1.3 · p. 66
• Later memories became more generalized and repetitive, while retaining procedural guidance about recipes, temperature measurement, state changes, circuits, and avoiding premature or looping actions.Source: §J.1.3 · pp. 66–67
• -8.66 average test score at round 20 followed 7.45 at round 1, while memory-bank size increased from 16 to 17 over 20 consolidation rounds.The experiment used 485 rollouts and examined initial, intermediate, and final memory banks.Source: §J.1.3 · pp. 65, 67

K.1. Auto-Mode Component Ablation

• Episodic Only recovers nearly all of the Auto gain, whereas Abstract Only never improves on zero-shot across four evaluation checkpoints.The ablation restricts access to distilled abstract memory, raw episodic memory, or both sources.Source: §K.1 · p. 68

K. Additional Analyses

• The Auto + Episodic evaluation couples abstract and raw memory sources, motivating separate component evaluation to identify their contributions.Source: §K.1 · p. 68

K.2. Memory Composition under Task Switching

• Fresh consolidation limits memory failures under task switching, while Cumulative consolidation buries the store in over-generalized and garbage entries.Fresh consolidates only the current task; Cumulative consolidates over tasks 1 through t.Source: §K.2 · p. 68, Figure 18 · p. 69
• Episodic Only accounts for nearly all of Auto’s gain, whereas Abstract Only stays at or below the zero-shot baseline throughout.Combining episodic and abstract memory adds a few points at each checkpoint, but the late decline at 400 follows erosion of the abstract store.Source: Figure 17 · p. 69
• Cumulative consolidation accumulates over-generalized and garbage memories that Fresh never accrues, with the gap widening across 15 ScienceWorld switch tasks and explaining a +203-point score gap.An LLM judge labeled each stored entry after every task; Cumulative’s failure totals diverged almost immediately and never recovered.Source: §K.2 · p. 68, Figure 18 · p. 69