Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents

Paper 2 (MRAgent) addresses a more fundamental and broadly applicable challenge—how LLM agents reason over long interaction histories—with a novel cognitive-science-inspired framework (associative memory graphs with active reconstruction). It demonstrates strong empirical gains (up to 23% improvement) on established benchmarks while reducing costs. Paper 1 (CMTF) solves a more specific problem (tool selection filtering) with practical but narrower impact. Paper 2's contribution to memory architectures has broader implications across diverse agent applications, making it more likely to influence future research directions in the rapidly growing LLM agent field.

Paper 1 addresses the notoriously difficult credit assignment problem in multi-agent systems using mathematically rigorous Shapley values. This represents a foundational algorithmic advancement with broad implications for training complex LLM agent architectures, whereas Paper 2 offers an architectural improvement to memory retrieval which, while highly relevant, is slightly less foundational in nature.

Paper 2 introduces a novel memory architecture (Cue-Tag-Content graph) plus an active reconstruction mechanism that tightly couples reasoning and memory access, addressing a central limitation of retrieve-then-reason pipelines. It reports sizable benchmark gains (up to 23%) alongside token/runtime reductions, suggesting strong methodological payoff and practical deployability for long-horizon agents. Its contribution is broadly applicable across agentic systems, retrieval-augmented reasoning, and efficient inference. Paper 1 is timely and important for safety evaluation of memory use, but is primarily an evaluation framework/measurement study with narrower technical generalization than a new memory paradigm.

Paper 2 likely has higher scientific impact due to strong real-world applicability (real-time autonomous driving), broad validation across many major datasets, and an end-to-end methodology that ties prediction to downstream safety via reinforcement learning—addressing deployment constraints and safety-critical objectives. Its heterogeneous distillation + PPO refinement is timely and practically actionable for edge deployment, with clear measurable gains (latency, compression, safety metrics). Paper 1 is novel for LLM-agent memory reconstruction and shows sizable benchmark improvements, but its impact may be more bounded to agent-memory research and depends on rapidly evolving LLM tooling and benchmarks.

Paper 2 likely has higher impact: it introduces a more conceptually novel memory paradigm (active reconstruction over an associative graph) that directly targets a widely recognized bottleneck—long-horizon agent memory/reasoning—while also improving efficiency (token/runtime), which is critical for deployment. The approach is broadly applicable across agentic tasks and architectures, and aligns with timely interest in scalable, adaptive memory systems. Paper 1 is strong and practical, but skill distillation from trajectories is closer to incremental advances in post-hoc tool/skill refinement and may be more domain/workflow-dependent.

Paper 2 addresses a critical bottleneck in the highly active field of LLM agents: long-context reasoning and memory retrieval. By shifting from a static retrieve-then-reason paradigm to dynamic, graph-based active reconstruction, it offers a highly novel and scalable solution. Its demonstrated performance improvements and cost reductions on standard benchmarks suggest broad, immediate real-world applicability across various autonomous agent tasks, giving it a higher potential for widespread scientific and practical impact compared to the architectural framework proposed in Paper 1.

Paper 2 offers a more broadly applicable and conceptually novel contribution: a reconstructed (active) memory access paradigm with an explicit Cue-Tag-Content graph that tightly couples retrieval and reasoning. This targets a central bottleneck for LLM agents across many domains (assistants, tools, robotics, scientific workflows), with demonstrated accuracy and efficiency gains on established long-context benchmarks. Paper 1 is impactful within AutoML/algorithm discovery, but its innovations are more system-integration and benchmark-specific, with narrower cross-field reach and potentially faster obsolescence as agent frameworks evolve.

Paper 2 has higher potential impact: it introduces a standards-derived, testable rubric that directly connects XAI outputs to evidence requirements in autonomous-driving safety assurance, addressing a timely and high-stakes deployment bottleneck. The approach is novel in framing “admissibility” via explicit lifecycle-stage criteria grounded in ISO/AMLAS, offering broad applicability across safety-critical ML domains and influencing both research and regulatory/industrial practice. While Paper 1 is technically innovative for LLM agent memory and shows strong benchmark gains, its impact is more confined to agent architectures and may iterate on an active retrieval trend rather than reshape evaluation/selection norms across fields.

Paper 2 has higher potential impact due to its timeliness and breadth: agent safety and reward hacking are central concerns for real-world deployment. It contributes a mechanistic monitoring framework that integrates internal activations with context/entropy to predict risky actions, evaluates in interactive environments (ALFWorld, WebShop), and explores mitigation (steering). This combination of measurement, causal-ish interpretation (latent policy state vs imminent action), and practical mitigation is likely to generalize across agent settings and influence both safety research and applied deployments. Paper 1 is useful for long-horizon memory, but narrower in cross-field implications.

Paper 2 is likely higher impact: it introduces a broad, large-scale benchmark (ConservationBench) and a systematic evaluation across 112 VLMs, delivering a strong, general negative result about physical transformation reasoning. This is timely and widely relevant to embodied AI, multimodal reasoning, and safety/robustness, and can influence model design and evaluation standards across the field. Paper 1 is innovative and practical for LLM agents, but its impact is narrower (agent memory architecture) and more incremental relative to existing retrieval/memory work.

Paper 2 likely has higher scientific impact: it introduces a technically novel agent-memory framework (graph-structured memory plus active reconstruction tightly coupled to reasoning), shows large empirical gains on established long-context benchmarks, and improves efficiency—key for real-world agent deployment. Its method is broadly applicable across LLM agents, retrieval/memory systems, and interactive tasks, making cross-field impact more likely and timely given current focus on long-horizon agents. Paper 1 is valuable for AI education assessment, but its impact is narrower and depends more on downstream adoption and human-validation work.

Paper 1 likely has higher impact: it is timely and highly relevant to real-world deployment of AI coding agents, introduces a new human-centric threat model, and provides rare large-scale empirical evidence (multi-hour tasks, >100 participants) showing severe oversight failure and partial ineffectiveness of monitors. The findings have immediate implications for software security, HCI, AI safety, and governance. Paper 2 is technically novel and useful, but memory architectures for LLM agents are a crowded area and results, while strong, are more incremental and narrower in cross-field societal urgency.

Paper 1 introduces a new LM architecture (ProtoT) replacing self-attention with a linear-cost prototype mechanism that yields concept-like internal representations, directly targeting interpretability-by-design while also improving scaling and robustness. This is a more foundational contribution with broader cross-field impact (architecture, efficiency, interpretability, safety) and strong timeliness given attention’s compute bottleneck. Paper 2 is impactful for agent memory and shows strong benchmark gains, but is more of a systems/framework advance within LLM agents and may generalize less broadly than a core architectural alternative.

RedKnot addresses a fundamental infrastructure bottleneck (KV cache management) affecting all long-context LLM serving, with broad applicability across the entire LLM ecosystem. Its head-aware decomposition is a novel architectural contribution that unifies multiple previously separate problems (prefix compression, hot/cold separation, distributed placement) into one framework without retraining. While MRAgent offers meaningful improvements for memory-augmented agents on specific benchmarks, RedKnot's impact spans a wider range of applications and establishes foundational infrastructure that could benefit the entire field of LLM deployment at scale.

Paper 2 (MRAgent) introduces a more broadly applicable and novel conceptual framework—active memory reconstruction inspired by cognitive science—that addresses a fundamental challenge (long-horizon memory reasoning) relevant across virtually all LLM agent applications. Its associative graph memory with dynamic reconstruction is a paradigm shift from static retrieve-then-reason approaches, with strong empirical gains (up to 23%). Paper 1 (MIRAGE) is well-executed but more narrowly focused on mobile UI agents and latent reasoning compression, which, while valuable, represents a more incremental advance in a specific application domain.

Paper 2 (MRAgent) addresses a broadly applicable problem—memory management for LLM agents—that impacts the entire AI community and numerous downstream applications. Its novel graph-based associative memory with active reconstruction is a generalizable architectural contribution with strong empirical results (23% improvement). Paper 1 (FeynmanBench) is a well-crafted benchmark but targets a narrow domain (Feynman diagrams in particle physics), limiting its breadth of impact. While it reveals important limitations of multimodal LLMs, benchmarks typically have less transformative impact than new architectural paradigms applicable across many domains.

Paper 2 addresses a fundamental and broadly applicable challenge—memory management for LLM agents across diverse tasks—with a novel biologically-inspired approach (reconstructive memory vs. retrieval). It demonstrates strong empirical gains (up to 23%) with practical efficiency benefits. Paper 1, while rigorous and interesting, targets a more specialized niche (optimization-style reasoning tasks for LLMs) with a narrower scope of applicability. Paper 2's graph-based associative memory framework has broader potential impact across many agent applications and connects to established cognitive science principles, giving it wider interdisciplinary appeal.