Product updates, engineering notes, and ideas on building agents that learn.
A nightly fine-tune and a one-line note appended to a markdown file are both sold as agents that learn from experience. They are the same loop with different parts swapped out. This is a map of the field: what changes, what decides, and when — and why the second question is the one that separates methods that compound from methods that only look like they do.
Dreaming makes an agent's memory better. Only measurement makes the agent better. Anthropic's Dreams, ChatGPT, Codex, Gemini CLI and OpenClaw all rewrite agent memory offline — and every gate they document checks whether the rewritten memory is tidy, not whether the agent improved.
Creao used Reflexio to turn successful agent behavior into reusable guidance—cutting median model calls by 75%, reducing task failures by 36%, and showing positive impact from retrieved learnings in 47% of evaluated cases.
AI memory and agent self-improvement sit in the same slot, use the same storage, and are sold against the same failure — which is why they get confused. The difference is what has to happen before the system writes anything down. One writes when something is said. The other writes when something is judged.
Loop engineering is how we get agents to iterate toward a goal on their own. Self-improvement is the next stage — a loop wrapped around those loops, so each pass makes the next one better. This is a field note on how the two ideas connect.
Memory remembers what happened. Learning changes what the agent does next time. LGRO — Learn, Generalize, Reflect, Optimize — is a practical framework for non-parametric, self-improving AI agents.