Developer Builds Coding Agent That Derives Rules From Past Corrections

A developer has built a coding assistant that moves beyond simple conversation recall by deriving standing rules from past corrections — a process the author calls 'reflect'. When told once to use date-fns instead of moment.js, the agent extracted and stored that preference as a project-wide convention, applying it automatically to future unrelated requests. The system works by injecting recalled memory context directly into the LLM's system prompt, requiring no separate teaching mode from the user. The author argues that most agent memory systems stop at retrieval, while the real value lies in a reasoning layer that generalises scattered corrections into durable rules. The approach is presented as broadly applicable beyond coding assistants, addressing any agent that repeatedly makes the same mistakes due to a lack of memory generalisation.