Memory Papers
Memory Architecture

CHIME: Credit-Aware Hierarchical Memory Evolution for Long-Horizon Agentic Planning

AuthorsYongshi Ye, Tian Lan, Feihu Jiang et al.

arXiv 20262026
arXivPDF

TL;DR

CHIME uses a Credit Attribution Gate over a Hierarchical Memory Bank to evolve stage-specific memories, reaching 39.01% average eval accuracy on DeepSeek-V4-Flash (+3.68 over A-MapReduce).

SharePost on XLinkedIn

Read our summary here, or open the publisher PDF on the next tab.

THE PROBLEM

Outcome-based memory causes biased planning feedback (credit assignment failure)

Self-evolving agents typically equate final task outcomes with plan quality, even though outcomes conflate planning, execution, and environment factors.

This mis-assigns credit in long-horizon agentic planning, so biased and noisy experience is repeatedly written into memory, causing systematic planning errors and contamination by low-level execution details.

HOW IT WORKS

Credit-Aware Hierarchical Memory Evolution

CHIME’s core mechanism is a Hierarchical Memory Bank with a Plan Bank, Execution Bank, a Credit Attribution Gate, and Credit-Aware Memory Evolution that updates only stage-attributed memories.

You can think of CHIME like a computer with separate “planning RAM” and “execution disk,” plus a judge that decides which component deserves each performance boost.

This attribute-before-memorize design lets CHIME store reusable, stage-specific guidance that a plain context window cannot track or selectively evolve over many long-horizon tasks.

DIAGRAM

One CHIME Episode: Planning and Execution Flow

This diagram shows how CHIME processes a single long-horizon task from query to trajectory using stage-specific retrieval and planning/execution.

DIAGRAM

Evaluation and Ablation Pipeline for CHIME

This diagram shows how CHIME is evaluated on train and eval splits and how ablations modify its components.

PROCESS

How CHIME Handles a Long-Horizon Agentic Task

  1. 01

    Hierarchical Memory Retrieval

    CHIME receives task xt, builds planning and execution queries, and retrieves Top-N memories from the Plan Bank and Execution Bank with similarity and value-based reranking.

  2. 02

    Planner and Executor

    Using retrieved planning memories, CHIME’s Planner generates a plan pt, then the Executor uses execution memories and pt to produce a trajectory τt for the task.

  3. 03

    Credit Attribution Gate

    The Credit Attribution Gate reviews xt, pt, τt, outcome st, and memories, assigning credit to planning, execution, both, or none, and generating stage-specific experience with confidence γt.

  4. 04

    Credit-Aware Memory Evolution

    CHIME’s Credit-Aware Memory Evolution updates memory values and contents only in the attributed banks, rewarding helpful items, penalizing misleading ones, and merging or inserting new experience.

KEY CONTRIBUTIONS

Key Contributions

  • 01

    Attribute-Before-Memorize Principle

    CHIME introduces an attribute-before-memorize pipeline with a Credit Attribution Gate and Credit-Aware Memory Evolution, achieving 34.87% eval average on Qwen3.5-Flash vs 29.50% for A-MapReduce (+5.37).

  • 02

    Hierarchical Memory Bank for Planning and Execution

    CHIME designs a Hierarchical Memory Bank with separate Plan Bank and Execution Bank, showing that removing either memory drops eval average by up to 4.81 percentage points.

  • 03

    Efficient and Transferable Memory Evolution

    CHIME accumulates only 129 memories on τ²-Bench versus 3,585 for A-MapReduce while maintaining higher accuracy, and its memory transfers across backbones with up to +4.68 percentage points over A-MapReduce.

RESULTS

By the Numbers

Average eval accuracy (DeepSeek-V4-Flash)

39.01%

+3.68 over A-MapReduce (35.33%)

Average eval accuracy (Qwen3.5-Flash)

34.87%

+5.37 over A-MapReduce (29.50%)

τ²-Bench eval (Qwen3.5-Flash)

25.92%

+5.37 over WebAnchor (20.55%) on train, +2.47 over WebAnchor (23.45%) on eval

BFCL-v4 eval (DeepSeek-V4-Flash)

67.15%

-2.83 vs A-MapReduce (69.98%) but +3.00 over WebAnchor (64.15%) baseline without memory

These numbers come from τ²-Bench, VitaBench, BrowseComp-ZH, and BFCL-v4, which test long-horizon customer service, life services, web information seeking, and function calling. The main result shows that CHIME’s credit-aware hierarchical memory yields higher average eval accuracy than A-MapReduce and other baselines on both Qwen3.5-Flash and DeepSeek-V4-Flash backbones.

BENCHMARK

By the Numbers

These numbers come from τ²-Bench, VitaBench, BrowseComp-ZH, and BFCL-v4, which test long-horizon customer service, life services, web information seeking, and function calling. The main result shows that CHIME’s credit-aware hierarchical memory yields higher average eval accuracy than A-MapReduce and other baselines on both Qwen3.5-Flash and DeepSeek-V4-Flash backbones.

BENCHMARK

Average Eval Accuracy on DeepSeek-V4-Flash Across Four Benchmarks

Average eval accuracy (%) comparing CHIME with baselines on DeepSeek-V4-Flash (Table 1).

BENCHMARK

Ablation: Eval Average on Qwen3.5-Flash

Eval average (%) for CHIME and ablations on Qwen3.5-Flash (Table 2).

KEY INSIGHT

The Counterintuitive Finding

CHIME achieves the highest accuracy on τ²-Bench while retaining only 129 memories, compared to 3,585 memories for A-MapReduce.

This is surprising because self-evolving agents are often assumed to benefit from larger memory banks, yet CHIME shows that quality and attribution, not quantity, drive long-horizon performance.

WHY IT MATTERS

What this unlocks for the field

CHIME unlocks stage-aware self-evolving memory where planning and execution experience are cleanly separated and updated with credit-aware signals.

Builders can now deploy long-horizon agents whose planning skills improve online without retraining, while avoiding contamination from execution errors and environment noise.

~13 min read← Back to papers

Related papers

Memory Architecture

A Control Architecture for Training-Free Memory Use

Yanzhen Lu, Muchen Jiang et al.

· 2026

TAG routes low-confidence steps to uncertainty-based routing, filters them with guarded acceptance with rollback, chooses between bank selection across rule and exemplar memory, and prunes via evidence-based retirement inside a unified control loop. On SVAMP and ASDiv, TAG reaches 81.0% and 85.2% accuracy, improving over the 74.0% and 77.5% no-memory baselines while a compute-matched Retry baseline stays flat.

arXiv:2604.18206Read explainer

Questions about this paper?

Paper: CHIME: Credit-Aware Hierarchical Memory Evolution for Long-Horizon Agentic Planning

Answers use this explainer on Memory Papers.

Checking…