AI coding agents can lose track of old project decisions
A common problem appears when an is expected to maintain through several files. The workflow uses files such as plan.md, , CONTEXT.md, and system_.md, with the AI agent updating them and the developer checking them by hand. Over time, those files can drift apart and start saying conflicting things.
When careful review slips, Cursor may ignore the planned system structure, forget features agreed on weeks earlier, and build the opposite of previous decisions. One proposed fix is an MCP-connected project canvas. The goal is to let the agent read live instead of relying on stale markdown files.
The item also asks s using Cursor, Claude Code, or similar tools to share their workflow through a short AI interview with no signup.
Key points
- Several AI-maintained can slowly fall out of sync.
- Conflicting project notes can make Cursor ignore earlier and feature decisions.
- The problem becomes more serious in long-running solo projects.
- The proposed fix is an MCP-connected project canvas that gives the agent live context.
- This is mainly workflow research, not a finished solution yet.