Ingest: Memory Intelligence Agent

Type: kb/sources/types/ingest-report.md

Source: memory-intelligence-agent.md Captured: 2026-04-11 From: https://arxiv.org/html/2604.04503v2

Classification

Type: scientific-paper — arXiv preprint with methodology, architecture, algorithm sketches, ablations, benchmark comparisons, and academic references.

Domains: agent-memory, trace-derived-learning, test-time-learning, reinforcement-learning

Author: Jingyang Qiao, Weicheng Meng, Yu Cheng, Zhihang Lin, Zhizhong Zhang, Xin Tan, Jingyu Gong, Kun Shao, and Yuan Xie — multi-institution team from East China Normal University, Shanghai Innovation Institute, Harbin Institute of Technology, Xiamen University, Shanghai AI Laboratory, and an independent researcher; worth attending to because the paper directly targets deep-research-agent memory and reports linked code/model/dataset artifacts.

Summary

MIA proposes a Manager-Planner-Executor architecture for deep research agents. The Memory Manager stores compressed historical search trajectories as explicit workflow memory; the Planner uses retrieved workflows to generate search plans and is updated through RL; the frozen Executor carries out tool-using research under the plan. The distinctive mechanism is a bidirectional loop between non-parametric and parametric memory during test-time learning: multiple plan rollouts produce tool/reasoning trajectories, an LLM Judger or unsupervised reviewer/area-chair process labels them, successful and failed trajectories are compressed into workflow memory, and the Planner parameters are updated from the same reward signal. The paper argues this avoids long-context memory bloat and improves planning, with ablations suggesting that memory-as-planning-prior and test-time learning matter more than simply feeding memory to the Executor.

Connections Found

The strongest connection is trace-derived-learning-techniques-in-related-systems: MIA extends that survey space as a mixed-substrate trajectory-derived learning system, because the same test-time trajectories promote into explicit workflow memory and Planner weights. It also extends the fundamental split in agent memory is not storage format but who decides what to remember by adding a design point where memory agency is distributed across Manager, Planner, Executor, Router, and Judger rather than held by one agent or one memory service. It exemplifies agent memory is a crosscutting concern: storage, retrieval/activation, learning, and action capacity are separate architectural roles. It extends memory management policy is learnable but oracle-dependent, since MIA learns a Planner policy from plan/trajectory rewards while trying to replace ground-truth labels with a weak unsupervised judge. Sibling sources include Trajectory-Informed Memory Generation, AgeMem, and A-MEM.

Extractable Value

  1. Mixed-substrate trace learning — MIA is a clean example where the same search trajectories become both inspectable-ish workflow memories and opaque Planner weight updates. This sharpens the substrate/timing matrix for continuous learning: trace-derived learning can run at test time and promote to artifacts, weights, or both. [deep-dive]

  2. Memory as a planning prior, not Executor context — the ablation reports that "Only Memory" can underperform, while "Memory for Planner" improves results. The transferable insight is that stored trajectories may be more useful for shaping a plan than for bloating the actor's direct context. [quick-win]

  3. Contrastive workflow memory — MIA selects the shortest successful trajectory and samples a failed trajectory, then stores them as positive/negative paradigms for future planning. This is a compact pattern worth comparing with ReasoningBank's success/failure extraction and trajectory-informed tips. [experiment]

  4. Weak-oracle design for open-world self-evolution — the unsupervised reviewer/area-chair judge decomposes evaluation into logic, format, and factuality reviewers plus a meta-decision. This is directly relevant to oracle theory: it is an attempt to turn one unreliable scalar judge into a structured bundle of weaker checks. [deep-dive]

  5. Frequency reward as retrieval diversity pressure — MIA retrieves memory by combining semantic similarity, value reward, and frequency reward. The frequency term explicitly rewards low-frequency memory units, a simple way to avoid always reusing the same high-similarity examples. [experiment]

  6. Planner/Executor stability split — the Executor is frozen during test-time learning while the Planner updates. This is a useful architectural pattern: keep the operational tool-using service stable while learning the strategy generator around it. [just-a-reference]

Limitations (our opinion)

Paper-only ingest; code not inspected. The snapshot links a GitHub repository, model, and dataset, but this ingest is based on the arXiv HTML text, not code review. Claims about what is implemented should stay at paper-confidence until the repository is reviewed separately.

Benchmark-bound oracle evidence. Most of the learning loop depends on correctness labels, LLM Judger outputs, or the proposed reviewer/area-chair surrogate. The paper reports benchmark improvements, but does not show that the unsupervised judge would remain reliable for open-ended research where there is no known answer and no clean factuality target.

Opaque policy updates inherit the usual inspectability costs. The Planner's learned improvements live in weights. That may be useful for speed and generalization, but the paper does not discuss rollback, debugging, provenance, or how to inspect a bad learned planning habit.

The simpler-account question remains open. MIA compares against several memory baselines and includes ablations, but the headline advantage could still partly come from extra rollout compute, any relevant plan examples, or a strong planner/router scaffold rather than the full bidirectional parametric/non-parametric memory loop.

No mature memory lifecycle. The Memory Manager stores compressed workflows and updates counts, but the paper does not offer a strong story for retiring stale workflows, detecting contradictory memories, or maintaining long-running memory quality beyond selective clearing after Planner training.

Deep-research specificity limits reach. The architecture is tuned around search trajectories, tool calls, and benchmarked question answering. Its strongest lessons transfer to agent learning loops, but the exact Manager-Planner-Executor setup may not transfer to KB curation, open-ended methodology work, or systems where the best output is not a final answer.

Update trace-derived-learning-techniques-in-related-systems.md: add MIA as a source-only mixed-substrate trajectory-run system. The note should say that MIA mines deep-research search trajectories into both structured workflow memory and Planner weight updates during test-time learning, and that its reviewer/area-chair unsupervised judge is a concrete weak-oracle attempt worth comparing against the survey's oracle bottleneck claim.