EvoWeave: Composing Reusable Agent Experience through Operation Level Graphs
EvoWeave addresses how interactive agents can reuse past experience without importing entire workflows or discarding step dependencies. It proposes a graph memory where nodes are abstract states and edges are reusable operations annotated with contracts, provenance, and capability references; a task planner guides retrieval and bounded graph search to assemble experience fragments. Experiments across four benchmarks and three LLM backbones show heterogeneous gains and highlight tradeoffs between retrieval precision and runtime overhead.
The paper presents a well-motivated framework for compositional experience reuse. The core abstraction—operations as directed state-transition edges with explicit input-output contracts—is a principled advance over flat or whole-workflow retrieval. EvoWeave achieves the highest reported mean in 16 of 21 backbone-metric combinations, and AppWorld ablations specifically attribute gains to directed connections. However, margins over strong baselines are often slim, the authors correctly avoid significance claims, and efficiency improvements are nuanced: while injected memory tokens drop sharply, total input tokens fall by less than 1 percent and per-round latency increases.
The operation-level graph representation directly tackles the composition problem by preserving input requirements and historical transitions while keeping capability references narrow. The controlled online-memory construction protocol—starting empty and accumulating experience identically across all methods—is a rigorous experimental design choice that isolates representation effects. The AppWorld ablation cleanly separates the contributions of decomposition, directed connections, and active retrieval, and the authors are transparent about negative results, such as cases where ReasoningBank or ExpeL outperform EvoWeave.
EvoWeave reduces injected memory tokens dramatically, yet total input tokens are nearly identical to No-Memory and per-round latency rises, so context efficiency does not translate clearly into speed or cost savings. Performance gains are inconsistent across backbones: ReasoningBank surpasses EvoWeave on several AppWorld Challenge and WebArena metrics, while ExpeL leads on ALFWorld and ScienceWorld, implying that graph-based retrieval is not universally dominant.
The evaluation is purely descriptive—no significance tests or confidence intervals are offered—and relies entirely on re-implemented baselines evaluated under the authors' harness, making it impossible to disentangle methodological superiority from implementation variance.
The AppWorld ablation provides the strongest evidence, showing that directed connections add 2.08 overall TGC points beyond isolated operation cards and that active retrieval provides a further lift. However, the comparison landscape is limited: all baseline numbers are internal re-implementations under a continual-learning protocol, so differences may reflect prompt engineering or retrieval tuning rather than core algorithmic advantages. The category analysis is a strength—it reveals wide internal variance, such as AppWorld Challenge scores ranging from 17.01 on Gmail to 36.83 on Spotify—but it also dilutes the interpretability of headline aggregates.
The authors supply pseudocode, default hyperparameters, and detailed aggregation rules, which aids reproducibility. Yet no public code repository or dataset snapshot is provided in the paper. Exact reproduction is explicitly acknowledged to be blocked by stochastic decoding and uncontrolled API model checkpoints. Moreover, the reliance on re-implemented baselines and the online memory construction protocol—where test tasks are processed sequentially and earlier experience feeds into later tasks—introduces ordering effects that are not ablated. The reflection and merge mechanisms rely on additional LLM calls whose full prompts are not released, leaving part of the system underspecified.
Different tasks often require interactive agents to perform similar operations at intermediate steps. Retrieving an entire past workflow can introduce irrelevant steps, whereas retrieving isolated operations can discard the dependencies needed to reuse them. We introduce EvoWeave, a framework for experience memory that represents abstract states as nodes and reusable operations as directed edges with input and output contracts, provenance, and capability references. A planner decomposes the current task into a dependency graph; operation retrieval and bounded directed search then assemble relevant experience fragments. An execution agent can inspect these fragments and their referenced capabilities, while reflection separates concrete execution evidence from reusable abstractions. We study the framework across four benchmarks, five evaluation settings, and three language model backbones. Across backbones, AppWorld Normal and Challenge task goal completion average 51.91 and 27.99, respectively, exceeding the baseline without memory by 6.99 and 4.99 percentage points. An AppWorld ablation associates directed connections with an additional improvement of 2.08 points over independent operation cards. The results motivate composition of reusable operations while exposing tradeoffs in retrieval overhead and conditions for valid reuse.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.