COVE: Adaptive Memory-Parameter Coordination for LLM Self-Evolution
Original title:Learning What to Remember and What to Internalize in LLM Self-Evolution via Adaptive Memory-Parameter Coordination
AI Summary
The paper presents COVE, a unified framework for LLM-agent self-evolution that coordinates two learning channels: harness-based updates, such as editable memories and skills, and parameter-based learning that internalizes experience into model weights. COVE uses task-aware routing, stage-aware scheduling, and knowledge optimization to match task or knowledge types with the appropriate channel. The authors report evaluations across multiple task categories, claiming more robust and efficient improvement than single-channel evolution strategies in changing environments. The abstract does not provide detailed benchmarks, baselines, model sizes, or quantitative results.
Why it's worth reading
As deployed agents face changing tools, APIs, and user requirements, deciding what to store externally versus internalize in parameters is becoming a central design problem for reliable self-improvement.
Deep Read
What happened
Original fact: An arXiv paper published on August 2, 2026, titled Learning What to Remember and What to Internalize in LLM Self-Evolution via Adaptive Memory-Parameter Coordination, introduces COVE, a framework for LLM-agent self-evolution. It targets deployed agents operating with changing tools, APIs, and user requirements.
Core technology
Original fact: COVE combines two learning channels. Harness-based learning writes feedback into editable memories or skills, while parameter-based learning internalizes experience into model parameters. The framework uses task-aware routing, stage-aware scheduling, and knowledge optimization to coordinate these channels.
Analysis: The central design treats the destination of experience as an adaptive decision: some knowledge should remain externally editable, while other knowledge may be suitable for longer-term parameter updates.
Key evidence & numbers
Original fact: The abstract states that COVE outperforms single-channel evolution strategies across multiple task categories and provides more robust and efficient improvement in changing environments.
Unverified from the supplied material: The abstract does not report task counts, model sizes, benchmark names, numerical gains, compute costs, routing accuracy, or statistical significance. The magnitude of the claimed advantage therefore cannot yet be assessed.
Why it matters
Analysis: External memory can be edited, rolled back, and updated quickly, but may introduce retrieval noise, context growth, and conflicting knowledge. Parameter learning can produce more persistent capabilities, but may incur training costs and risks such as catastrophic forgetting or internalizing incorrect experience. Coordinating both channels could provide a finer control layer for deployed agents.
Practical impact
Analysis: If confirmed by the full experiments, agent builders could keep volatile, environment-specific, or user-specific information in memories and skills, while directing stable and transferable capabilities toward parameter learning. Production systems would still need versioning, experience filtering, rollback, and safety evaluation, especially for tool-use workflows.
Limitations & uncertainty
Original fact: The supplied information is limited to the abstract. It does not specify the router implementation, scheduling frequency, parameter-update method, knowledge-optimization objective, or whether code is available.
Analysis: Claims of better robustness and efficiency may depend strongly on task construction, baseline selection, and compute budgets. Without long-horizon environment changes, transfer to unseen tasks, memory-contamination tests, and explicit parameter-update cost comparisons, deployment relevance remains uncertain.
Original sources
- arXiv abstract page
- Paper: Learning What to Remember and What to Internalize in LLM Self-Evolution via Adaptive Memory-Parameter Coordination
- Identifier: arXiv:2608.01234