Decision-Aware Memory Cards improve tool-using LLM agents

The arXiv paper "Decision-Aware Memory Cards: Counterfactual-Inspired Context Selection and Compression for Tool-Using LLM Agents" was submitted on 6 June 2026 by Xinyu Guan and two coauthors, per arXiv. The paper proposes CICL, a "decision-aware context layer" that turns instance evidence into a context graph and produces typed memory cards for budgeted agents, according to the arXiv abstract. The submission reports empirical gains on a retrieval task and includes controlled diagnostics and auxiliary agreement checks, per arXiv.

Per the arXiv abstract, CICL routes deterministic and model-assisted judgments from models including Opus-assisted, Qwen, Codex/GPT-5.5, and Qwen-QLoRA through a shared eight-field schema, scores candidate evidence units on metrics labeled action shift, outcome uplift, necessity, and negative-transfer risk, and packs high-utility evidence into compact "memory cards" for an agent with a context budget. The paper reports that Qwen3.6-plus reranking of BM25 top-50 candidates on 50 SWE-bench Verified file-retrieval instances raises hit@1 from 0.58 to 0.78 and MRR@10 from 0.634 to 0.790, with all 2,500 judgments parseable. Controlled diagnostics show at budget 120, CICL reaches F1 0.620 on v1 and 0.425 on v3, and removing the top-utility semantic v3 unit collapses F1 to 0.000. The abstract also notes smaller agreement checks using Qwen-QLoRA, a 200-label Opus-assisted signal, and a three-instance patch smoke validating retrieval-to-patch plumbing, and states that RepoBench-R summaries still outperform the cards and that compact rankers do not yet replace the heuristic, per arXiv.

Decision-aware selection addresses a common failure mode in retrieval-augmented agents where decisive evidence exists but is not surfaced or compressed at action time. Industry-pattern observations indicate scoring candidates by downstream decision impact, rather than raw similarity alone, can materially change which evidence guides actions in tool-using agents. The reported jump in hit@1 and MRR suggests that reranking with a decision-focused signal can improve retrieval precision on task-specific benchmarks, though the paper's own caveats about limits and comparisons to summary baselines temper broad generalization.

reproducibility of the reported metrics on larger, diverse benchmarks; head-to-head comparisons with dense retrieval and learned rankers; latency and compute costs of the judgment layer; and how decision-aware selection composes with retrieval-augmented generation pipelines. The arXiv submission frames CICL as a measurement and selection layer rather than a turnkey agent fix, per arXiv.