Long context helps coding agents, but it is not real memory
can feel much stronger in coding work because it can handle a very . It can plan better, notice edge cases, and stay more stable during longer architecture tasks. But a does not automatically become useful memory.
In one workflow, an agent inside an tried to install a package, ran into a dependency problem, made several bad attempts, and eventually found the working fix. Later, when a similar problem appeared, the agent did not cleanly recall only the final fix. It pulled in the whole messy history, including failed commands and temporary workarounds.
Broken install attempts started to look like useful clues, and one short-term assumption from that specific container began showing up in other environments. A stores more history, but it does not decide what was wrong, what was temporary, and what should be reused.
Key points
- felt stronger on longer coding and architecture tasks.
- A large context can keep too much messy history, including failed attempts.
- The agent later reused broken install attempts as if they were helpful.
- A workaround from one appeared in unrelated environments.
- s should separate final fixes from failed experiments when guiding a .