💬 NLP

MemoHarness: Agent Harnesses That Learn from Experience

MemoHarness is an adaptive agent optimization framework that enhances LLM performance across diverse tasks by decomposing the control layer into six dimensions and dynamically tailoring harness configurations for each case using a dual-layer experience bank derived from past executions, all without requiring test-time labels or additional search.

Original authors: Yue Huang, Wenjie Wang, Han Bao, Yuchen Ma, Xiaonan Luo, Yi Nian, Haomin Zhuang, Zheyuan Liu, Yue Zhao, Xiangliang Zhang

Published 2026-07-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yue Huang, Wenjie Wang, Han Bao, Yuchen Ma, Xiaonan Luo, Yi Nian, Haomin Zhuang, Zheyuan Liu, Yue Zhao, Xiangliang Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a brilliant, super-smart robot assistant. You give it a task, and it tries to solve it. But here's the catch: the robot doesn't just need to be smart; it needs to know how to use its brain. It needs a set of instructions on how to organize its thoughts, when to look up information, how to ask for help, and how to double-check its work before giving you the answer. In the world of artificial intelligence, this set of instructions and tools is called an "agent harness." Think of it like the dashboard, steering wheel, and GPS in a car. The engine (the AI model) might be powerful, but if the dashboard is broken or the GPS is stuck on an old map, the car isn't going to get you where you need to go. For a long time, scientists have been trying to make these robots smarter by making the engines bigger. But this paper asks a different question: what if we just fixed the dashboard? What if we could teach the robot to learn from its own mistakes and adjust its driving style for every single trip, rather than using the same stiff, one-size-fits-all instructions for everything?

This is exactly what the researchers behind MemoHarness set out to do. They realized that most AI agents today are like cars with a fixed steering wheel: they use the same setup for every problem, whether it's a simple turn or a complex highway merge. The team built a new system that acts like a "learning driver." Instead of just searching for one perfect set of instructions to use forever, MemoHarness lets the agent learn from its past trips. It keeps a detailed diary of what worked, what failed, and why. When a new task comes up, the agent doesn't just guess; it looks at its diary, finds similar past situations, and tweaks its own instructions on the fly to fit the specific challenge.

The paper introduces a clever way to break down the "harness" into six distinct parts, like adjusting different knobs on a control panel: how the agent gathers information (Context), what tools it uses (Tool), how it thinks and speaks (Generation), the order of its steps (Orchestration), what it remembers (Memory), and how it finalizes its answer (Output). By treating these as separate, editable surfaces, the system can diagnose exactly what went wrong. Did it forget to check a map? Did it try to drive too fast? Did it forget to remember a key detail?

In their experiments, the team tested this system on three very different types of tasks: acting like a computer terminal to fix software, writing code, and solving complex financial reasoning puzzles. The results were promising. On the computer terminal tests, MemoHarness improved the success rate from 0.722 to 0.806, beating the best fixed setups they compared it against. It also showed that these learned habits could "transfer" to new, unseen tasks and even work well on different types of AI models without needing to be retrained. For instance, when they took the harness learned on one model and applied it to six other models, the average success rate jumped by +0.098.

The authors suggest that this approach is a practical way to make AI more adaptable without needing to build a new engine every time. However, they are careful to note that while the results are strong, they are based on specific tests and simulations. They don't claim to have solved every problem or proven that this works perfectly in every single real-world scenario. They also found that while the system uses more data to "read" its diary, most of that data can be cached (stored for quick reuse), keeping the cost competitive. In short, MemoHarness suggests that giving an AI the ability to reflect on its own experience and adjust its own control panel might be the key to making it truly helpful, turning a rigid machine into a flexible, learning partner.

Technical Summary: MemoHarness

Problem Statement

The performance of Large Language Model (LLM) agents is heavily dependent on the agent harness—the external control layer that manages context construction, tool access, inference orchestration, memory retention, and output validation. While practitioner experience indicates that harness design can swing task success by tens of percentage points, current automated optimization methods typically focus on narrow artifacts (e.g., prompts, pipelines, or workflows) rather than the full control layer. Furthermore, deployed agents usually rely on a single, static global harness for all cases, failing to adapt to specific variations in domain, ambiguity, or reasoning depth.

Existing approaches face three primary challenges:

High Dimensionality and Coupling: The harness is a complex system where changing one component (e.g., retrieval) affects others (e.g., prompt format or memory policy).
Weak Supervision: Benchmark scores indicate success or failure but do not diagnose which harness dimension caused the failure or what should be transferred to future cases.
Test-Time Adaptation Gap: Deployed agents lack the ability to specialize their harness for a specific incoming case without requiring new labels, feedback, or a fresh search run.

Methodology: MemoHarness

MemoHarness is an adaptive framework that learns from its own executions to optimize the agent harness. It operates in two phases: Training-Time Search and Test-Time Case Adaptation.

1. Six-Dimensional Harness Space

Instead of treating the harness as a monolithic prompt, MemoHarness decomposes it into six editable control dimensions based on the temporal flow of inference:

D1 Context Assembly: Construction of model input (instructions, constraints, retrieved material).
D2 Tool Interaction: Control over when and how external tools or retrievers are called.
D3 Generation Control: Decoding parameters (sampling, token budget).
D4 Orchestration: Workflow topology (sequence of calls and reasoning steps).
D5 Memory Management: State persistence and context pruning across calls.
D6 Output Processing: Post-processing and validation of raw model outputs.

2. Dual-Layer Experience Bank

During the search phase, MemoHarness accumulates a structured experience bank ( $B_t = \{E_t, G_t\}$ ):

Per-Case Entries ( $E_t$ ): Records execution trajectories, rewards, costs, and a diagnosis for each case. The diagnosis identifies the primary failure dimension, secondary contributing dimensions, and a natural language analysis.
Global Patterns ( $G_t$ ): Distilled from failure clusters, these patterns capture recurring phenomena, supporting evidence, and the expected effect of targeted harness changes.

3. Training-Time Optimization

The system performs a guided search over the six-dimensional space starting from a minimal harness.

Correctness-First Selection: The primary objective is maximizing task reward (correctness). Execution cost (token usage) serves only as a secondary tiebreaker.
Iterative Refinement: A controller proposes new harness configurations based on retrieved evidence from the experience bank. The search runs for $T$ iterations, accumulating diagnostic knowledge rather than just scalar scores.

4. Test-Time Case Adaptation

At inference time, the system does not perform a new search. Instead, it adapts the learned global harness ( $W^*$ ) to a specific test case ( $x_j$ ):

Retrieval: The system retrieves similar historical cases (both successes and failures) and relevant global patterns from the frozen experience bank based on case features and instruction similarity.
Adaptation: A controller maps the global harness to a case-specific harness ( $W(x_j)$ ) using the retrieved evidence.
Execution: The adapted harness is executed to produce the final output. No test-time feedback or gradient updates are used.

Key Contributions

Structured Harness Optimization: A framework that decomposes the monolithic harness into six separable control surfaces paired with a dual-layer experience bank, enabling search to accumulate reusable diagnostic knowledge.
Test-Time Adaptation Mechanism: A method to adapt a search-derived global harness to individual test cases using retrieved experience, operating without test-time labels or feedback.
Empirical Validation: Results demonstrating improved task success across in-domain, cross-dataset, and cross-model evaluations, along with favorable cost behavior when retrieved context is cacheable.

Experimental Results

The authors evaluated MemoHarness on three benchmarks: Terminal-Bench (shell tooling), LiveCodeBench (code generation), and FinanceAgent (analytical reasoning).

Task Success (RQ1): On Terminal-Bench, MemoHarness achieved a mean success rate of 0.806, outperforming the strongest fixed-harness baseline (Codex) at 0.722.
Search Evolution (RQ2): The harness quality improved over iterations, with the most significant gains observed on long-horizon tasks (FinanceAgent), where the base harness left substantial room for improvement.
Cross-Dataset Transfer (RQ3): Learned harnesses showed selective transfer to unseen evaluation suites. For instance, a harness learned on Terminal-Bench improved performance on MMMLU, StrongReject, and SWE-Bench Pro, though gains were not uniform across all suites.
Cross-Model Transfer (RQ4): A harness learned using GPT-5.3-Codex was applied to six other models without retraining. All models showed improvement, with a mean gain of +0.098 across the family, suggesting the learned control policies are portable.
Cost Efficiency (RQ5): While MemoHarness retrieves additional context, most of this input is cacheable. In the evaluation, MemoHarness achieved higher accuracy than Codex and Claude Code at a lower total dollar cost ($6.89 vs. $10.28 and $9.51, respectively), provided the caching assumption holds.

Significance and Claims

The paper posits that execution experience is a practical substrate for building agent harnesses that are more adaptive than a single static configuration. MemoHarness demonstrates that the control layer surrounding an LLM can be optimized to improve end-to-end performance, serving as a practical complement to model scaling and manual engineering.

The authors explicitly state that while their results provide evidence for the efficacy of adaptive harnesses, broader claims regarding statistical robustness and detailed component attribution are left for future work. They emphasize that the observed gains are selective rather than universal, and the framework relies on the assumption that retrieved experience can be effectively cached to maintain cost competitiveness.