Persistent memory lets corrections compound

A directory that logs corrections, read at session start, means every fix compounds across sessions. The model stops repeating old mistakes. This is the single biggest multiplier for daily-driver agents. The agent reads the memory before every task, starting with accumulated context (see: context management is the dominant variable in agent quality and cost) rather than a blank slate.

The mechanism is simple: a markdown file in the repo that records corrections, preferences, and project context. The agent reads it at the start of every session. When you correct the agent’s assumption about a naming convention, you add it to the file. Next session, the agent already knows. Over weeks, the file becomes a curated record of everything the agent got wrong and learned from. This turns one-time corrections into permanent improvements, which is the single highest-leverage habit for daily-driver agent use.