Most agents start every session from zero. They can reason, call tools and follow instructions, but they don’t get better at the job they were deployed to do. Continual learning has usually meant updating model weights, which is slow, expensive and hard to audit.

This session takes a different route: continual learning in token space. The agent improves because you engineer what it retains, how it refines that knowledge over time, and what it recalls into context when it matters. Richmond builds a memory-aware agent on a real memory substrate and works through the four Rs of memory engineering: retain, reuse, refine and recall.

You’ll see how to extract memory from execution traces, consolidate it and resolve conflicts, forget without deleting, and run evaluations that show whether the agent improved between sessions.