ViktorAxelsen/MemSkill

What it solves

MemSkill addresses the limitation of static, hand-designed memory operations in long-horizon agents. Instead of relying on fixed rules for what to remember, it enables agents to learn and evolve "memory skills"—meta-memory strategies that determine what information to extract, how to remember it, and what to preserve or forget based on task feedback.

How it works

MemSkill implements a data-driven loop consisting of a controller, an executor, and a designer:

• Skill-Conditioned Construction: The system uses a small set of relevant skills to construct memories in a single pass for each span of interaction.
• Skill Evolution: The framework periodically mines challenging "hard cases" from task failures to refine existing skills or propose new ones.
• Reusable Skill Bank: It maintains a shared bank of skills that can be transferred across different datasets and base LLM models.
• Training Loop: For interactive environments like ALFWorld, it uses offline expert trajectories for scalable memory construction training and environment rollouts for task-level reward feedback to optimize the controller policy via PPO.

Who it’s for

Researchers and developers building long-horizon AI agents that need to manage large amounts of interaction history and improve their memory efficiency and quality over time.

Highlights

• Meta-Memory Approach: Focuses on the skill of remembering rather than the content itself.
• Self-Evolving: Automatically refines memory strategies based on performance feedback.
• Transferable: Supports a reusable skill bank that works across different base models and datasets.
• Scalable Infrastructure: Includes multi-API-key round-robin and multi-processing for high-throughput evaluation and training.