XNews

Memento Open-Source Framework Enables LLM Agents to Learn Continually Without Weight Updates

Memento Open-Source Framework Enables LLM Agents to Learn Continually Without Weight Updates

Memento, a 100% open-source memory-based continual learning framework, enables LLM agents to learn from experience over time without modifying model weights. It maintains a Case Bank of past trajectories and uses a planner-executor architecture with Model Context Protocol tools to retrieve and reuse similar cases for new tasks, reporting strong results on long-horizon and out-of-distribution benchmarks.

Event version 1.

Key insights

Learning without weight updates

Memento lets LLM agents improve through stored experience rather than fine-tuning model weights. It retrieves and reuses relevant past cases when handling new tasks.

Case Bank of past trajectories

The framework maintains a Case Bank recording tasks, step sequences, tool usage, and outcomes. New tasks are planned by pulling from similar past cases instead of starting from zero.

Planner and executor architecture

An LLM-based Planner decomposes tasks into subtasks, retrieves relevant cases, and selects a plan. An Executor runs subtasks through tools such as search, code execution, and document processing via the Model Context Protocol, then logs results back into memory.

Unified tool ecosystem

Memento provides a unified tool ecosystem covering search, code execution, document processing, and media analysis. The framework is 100% open source.

Reported benchmark performance

Memento reports strong results on long-horizon and out-of-distribution tasks in reported benchmarks. The material does not name the benchmarks or provide quantitative figures.

Source posts