MemRL: Self-Evolving Agents via Runtime Reinforcement Learning on Episodic Memory
What it is
MemRL is a Python library and research implementation for a non-parametric agent framework that improves performance through reinforcement learning on episodic memory. It uses a Two-Phase Retrieval mechanism to filter noise from past experiences and update strategies based on environmental feedback without updating model weights.
The problem it solves
Current AI agents struggle to self-evolve; fine-tuning is expensive and causes catastrophic forgetting, while existing memory methods rely on passive semantic matching that retrieves noise.
What you could use it for
- Improving agent performance on coding benchmarks like BigCodeBench
- Enhancing reasoning capabilities on Humanity's Last Exam (HLE)
- Optimizing task completion in interactive environments like ALFWorld
- Lifelong learning in database and OS interaction tasks via Lifelong Agent Bench
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.