Source-linked AI summary
When Successful Memories Mislead Embodied Agents:Memory Adaption For Task-Conditioned Execution
Quanquan Li, Hongbo Zhang, Yihe Chi, Liuyang Song, Jingyu Li, Yuxiang Huang, Hongzhen Zhang, Guitao Cao
TL;DR
Previously successful, relevant memories may still contain actions or state assumptions unsuitable for the current task, and a uniform representation may not preserve what each execution regime needs. MATE deterministically adapts retrieved trajectories to the task and budget without additional LLM inference. On ALFWorld, it recovers performance lost under direct trajectory reuse, with action normalization identified as the principal supported mechanism.
Problem
Retrieved trajectories can be semantically relevant and historically successful yet contain actions or state assumptions incompatible with the current execution context.
Source: §1 · p. 1Method
MATE deterministically adapts retrieved trajectories by filtering obsolete context, extracting transitions, normalizing actions, and rendering task-conditioned memory under a budget without additional LLM inference.
Source: §1 · p. 1Results
Takeaways & Limitations
Controlled comparisons identify action normalization as the principal supported mechanism, while fresh-memory and planning results support selective rather than universal adaptation.
Source: §1 · p. 1Safe transformations remain challenging when action specifications are incomplete or retrieved actions fall outside the verified mapping.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Experience reuse can reduce repeated exploration in embodied agents, but a trajectory that succeeded previously may be unsuitable for the current execution context. Existing memory systems pri marily optimize construction and retrieval; semantic relevance and historical success therefore remain insufficient when retrieved ex perience contains incompatible actions or an inappropriate level of structure. We introduce Memory Adaptation for Task-Conditioned Execution (MATE), a deterministic post-retrieval procedure that converts trajectories into execution-oriented memory. MATE re moves obsolete control context, extracts condition-action-effect transitions, applies verified action normalization, selects a task dependent representation, and serializes the result under a fixed budget without additional LLM inference. On 134 ALFWorld tasks, MATE achieves task success rates of 81.3% and 93.3% with Qwen2.5-14B and 72B while using approximately one-tenth of the tokens required by raw trajectories. Controlled comparisons show that verified action normalization is the principal mechanism by which MATE restores the utility of retrieved experience, support ing memory adaptation as a distinct stage between retrieval and embodied execution.
1. INTRODUCTION
Previously successful, relevant trajectories can still mismatch the current state or action space, while the needed memory structure varies by task. MATE adapts retrieved experience before execution, and controlled ALFWorld results identify action normalization as its principal supported mechanism.
1. INTRODUCTION
- A trajectory successful in its source episode can induce invalid actions and reduce performance when current objects, states, or action conventions differ.Semantic relevance and historical success are therefore insufficient to establish current suitability.Source: §1 · p. 1
- Separating excessive context, irrelevant content, and action incompatibility requires fixed retrieval, paired outcomes, and controls that manipulate memory length and action compatibility independently.The paper uses a factorized evaluation protocol to make transformations operationally explicit and causally testable.Source: §1 · p. 1
- Task structure changes what memory must preserve: reactive control depends on local transitions, whereas long-horizon planning needs dependencies spanning decisions.A uniform compressed representation may lose task-dependent information.Source: §1 · p. 1
- MATE deterministically filters obsolete control context, extracts transitions, normalizes actions, renders memory for the task profile, and serializes it under a budget without additional LLM inference.Reactive control receives local records; planning receives global constraints.Source: §1 · p. 1
- MATE recovers performance lost under direct trajectory reuse and outperforms a budget-capped, action-normalized full-trajectory control, while remaining statistically comparable to source-matched actions.The introduction identifies action normalization as the principal supported mechanism.Source: §1 · pp. 1–2
- The introduction frames MATE as post-retrieval adaptation that addresses unsuitable action forms and evaluates when adapted memory is preferable to other representations.The stated contributions include diagnosing negative transfer and delimiting when action sequences or full trajectories remain competitive.Source: §1 · pp. 1–2
2. RELATED WORK
Prior work has reused and structured embodied-agent experience, but selected memories are generally passed to the executor without an explicit check of action and state compatibility. Negative-transfer findings and task-dependent representation needs motivate deterministic post-retrieval adaptation.
2.1. Memory for Embodied and Interactive Agents
- Embodied and interactive agents reuse memory across episodes; prior approaches used verbal feedback or distilled experience and structured histories as workflows, manuals, replay records, procedural memories, or hierarchical working memory [8] [10] [11].These representations improve construction and retrieval but generally do not explicitly test whether action forms and state assumptions remain valid for the current task.Source: §2.1 · p. 2
2.2. Experience Reuse and Negative Transfer
- Experience reuse can produce negative transfer: prior work reports error propagation, conflicting advice, spurious trajectory correlations, and mismatches between transfer conditions and retrieval organization [19] [22] [20].Candidate memories may require verification before execution [21], and semantic relevance or historical success does not guarantee current utility.Source: §2.2 · p. 2
2.3. Task-Conditioned Trajectory Representations
- Local condition-action-effect records are economical for reactive control but may omit cross-step dependencies, whereas full trajectories retain global context alongside irrelevant or invalid actions.This tradeoff motivates normalizing historical actions and rendering memory according to task structure rather than applying a uniform summary.Source: §2.3 · p. 2
3. METHODS
MATE is a deterministic post-retrieval procedure that transforms retrieved trajectories into execution-oriented memory without changing retrieval or the executor. It preserves task-relevant transitions or constraints while applying the current action specification and token budget.
3.1. Embodied Task and Adaptation Objective
- MATE separates memory adaptation from retrieval quality and task-fact correctness by targeting applicable transition or constraint structure under the current action specification and budget.It addresses cases where historical success does not imply current applicability because object states, action syntax, or planning structure may differ.Source: §3.1 · p. 2
3.2. Deterministic Memory Adaptation Protocol
- The deterministic protocol filters context, extracts execution structure, normalizes supported actions, renders a task-conditioned representation, and serializes it within the token budget.Only exact rule matches are rewritten; identical inputs yield identical serialized memory, and shortening context alone does not enforce action validity or task-dependent structure.Source: §3.2 · p. 2
3.3. Transition Extraction and Action Normalization
- For message-form trajectories, MATE removes the historical instruction and following acknowledgement while retaining the interaction trace, then pairs explicit actions with nearby environment messages.Serialized traces use the preceding observed effect as the next precondition; free-form thoughts are not copied into the procedural record.Source: §3.3 · p. 2
- Verified action normalization rewrites only exact supported source forms, leaving unmatched expressions unchanged and auditable.MATE does not infer unseen action semantics.Source: §3.3 · p. 2
3.4. Task-Conditioned Rendering and Serialization
- The supplied task profile selects local precondition-action-effect records for reactive control or current-context global constraints for planning.Serialization follows retrieval order, fits complete items where possible, and records source, selection, token, normalization, and hash metadata for traceability.Source: §3.4 · p. 3
4.1. Datasets and Benchmarks
The evaluation includes embodied ALFWorld and ScienceWorld benchmarks, plus a non-embodied TravelPlanner boundary test.
4.1. Datasets and Benchmarks
- Evaluation covered 134 ALFWorld tasks with a separate pool of 252 successful training trajectories, 149 ScienceWorld tasks, and 1,000 strict-Final-Pass TravelPlanner instances.TravelPlanner served as a non-embodied boundary test.Source: §4.1 · p. 3
4.2. Implementation Details
Implementation used a ReAct loop and fixed evaluation settings, with MATE inserted after retrieval.
4.2. Implementation Details
- The ReAct loop used Qwen2.5-14B-Instruct, Qwen2.5-72B-Instruct, and gpt-5.4-mini, with decoding temperature 0, seed 0, and at most 30 steps.Qwen runs fixed the BGE-M3 retriever, queries, k = 10, pool, and manifests.Source: §4.2 · p. 3
- Historical ALFWorld trajectories used the legacy action specification, while execution used the current specification; a frozen deterministic map normalized supported legacy actions.Unmatched expressions remained unchanged.Source: §4.2 · p. 3
- MATE was inserted only after retrieval while queries, pool, retriever, and harness were reused; aggregate-only published-baseline comparisons are descriptive.Exact McNemar tests require task-level paired outcomes.Source: §4.2 · p. 3
4.3. Controls and Evaluation Metrics
Controls used frozen retrieval to compare memory representations under a shared token ceiling, with paired tests and separately corrected comparison families.
4.3. Controls and Evaluation Metrics
- Under a 2,048-token ceiling, raw prefix-filled memories averaged 2,004 tokens, while capped normalized-full memories averaged 1,501/1,500 tokens at 14B/72B.The shared ceiling yielded different packing and realized lengths; the action-sequence control used MATE's selected identifiers and ranks.Source: §4.3 · p. 3
- Evaluation reported TravelPlanner Final Pass Rate and used two-sided exact McNemar tests for paired outcomes.Holm correction was applied separately within diagnostic, published-method, representation, and 14B ablation families.Source: §4.3 · p. 4
- Boundary tests report unadjusted p-values, while reported pH values are adjusted within each comparison family.The adjustment scheme separates diagnostic, published-method, representation, and 14B ablation comparisons.Source: §4.3 · p. 4
4.4. Experimental Results and Analysis
The experiments compare MATE’s main results, mechanism controls, ablations, and robustness boundaries. They find strong gains over raw trajectory reuse, with normalization central to the supported benefit and adaptation not uniformly preferable across tasks.
4.4.1. Main Results
- MATE recovered 61.2 and 23.9 percentage points over raw full trajectories, with paired wins/losses of 85:3 and 34:2, and had the highest SR among single-pass methods at both scales.Three-attempt Reflexion did not differ significantly from single-pass MATE.Source: Table 1 · p. 4, §4.4.1 · p. 4
4.4.2. Mechanism Analysis
- Normalizing 19 of 877 lines raised SR to 82.1% and 88.1%, whereas prefix-filled raw trajectories gained only 3.0 and 11.2 points.MATE used 10.5% of raw-full tokens and reached 81.3% and 93.3%; uncapped normalized full was 0.8 points higher at 14B.Source: Table 2 · p. 4, §4.4.2 · p. 4
4.4.3. Ablation and Budget Control
- Removing normalization reduced 14B SR from 81.3% to 51.5%; retaining obsolete context yielded 73.1%.MATE exceeded the capped normalized-full control, while comparison with source-matched actions found no significant difference; richer transition fields had no established independent accuracy gain.Source: Table 2 · p. 4, §4.4.3 · p. 4
- Removing the budget preserved 14B SR at 81.3% but increased mean length from 1,617 to 7,861 tokens, making the observed effect one of efficiency rather than accuracy.Against source-matched actions, MATE-only/control-only/tied outcomes were 15/12/107 and 5/3/126.Source: Table 2 · p. 4, §4.4.3 · p. 4
4.4.4. Robustness and Boundary Analysis
- With gpt-5.4-mini, raw full trajectories reduced SR from 78.4% to 59.7%, extending memory-induced negative transfer beyond the Qwen executors.The comparison recorded 38 harms and 13 rescues (praw = 6.2 × 10−4).Source: Table 3 · p. 4, §4.4.4 · p. 4
- Full trajectories and MATE were comparable with fresh ALFWorld memory and in ScienceWorld, while TravelPlanner’s full trajectories led MATE by up to 4.9 points.On TravelPlanner, global rendering exceeded local cards only directionally at 14B (0.2% versus 0.0% FPR), supporting selective rather than universal adaptation.Source: Table 3 · p. 4, §4.4.4 · p. 4
5. DISCUSSION
MATE separates retrieval from the choice of what representation reaches the executor. The discussion retains raw trajectories as provenance while identifying conditions that complicate safe transformation.
5. DISCUSSION
- The proposed architecture assigns retrieval the role of finding potentially useful experience and adaptation the role of choosing what representation reaches the executor.Raw trajectories can remain as provenance while the system exposes an adapted local representation or fuller context for long-horizon planning.Source: §5 · p. 4
- The appropriate exposed representation can differ between local execution and long-horizon planning.The discussion contrasts adapted local representations with fuller context for planning.Source: §5 · p. 4
- Safe transformations remain an open challenge when action specifications are incomplete or retrieved actions fall outside the verified mapping.Source: §5 · p. 4
6. CONCLUSION
The conclusion presents MATE as deterministic post-retrieval adaptation for embodied agents. On ALFWorld, it reports the highest single-pass success rate at both model scales while substantially reducing memory length.
6. CONCLUSION
- MATE deterministically filters obsolete context, normalizes historical actions, and renders task-conditioned memory without additional LLM inference.Source: §6 · p. 4
- On ALFWorld, MATE achieved the highest single-pass SR at both model scales.Source: §6 · p. 4
- MATE substantially reduced memory length on ALFWorld.Source: §6 · p. 4