This article presents the MemCryst framework, which achieves lifelong context adaptation by bidirectional learning of memory and skills. Lifelong LLM agents must operate within a bounded context window while continuously expanding their interaction history. Memory makes past experience accessible, and skills offer reusable procedures, yet each alone may miss task‑specific facts. MemCryst composes the agent’s context from retrieved memory records and learned skills, and after each decision it updates the influence of those records based on task outcomes without altering the records themselves. The core loop consists of:
- Storing recurring episodes in a memory bank to create or revise skills;
- Guiding interactions with those skills, thereby generating new memory entries for future learning.
We evaluate across ten settings—embodied control, long‑document reading, conversation, mathematics, and code. In the ALFWorld benchmark, joint memory‑skill serving reaches 63.17% success, outperforming memory‑only (59.83%) and skill‑only (50.50%). The implementation is open‑source:
https://github.com/tianyi0216/Memcryst.git
Review: MemCryst demonstrates the synergistic benefits of coupling memory and skills, offering a scalable route for continual learning in LLM agents.