How do you get an AI agent back on track when it loses context in a long chat?
Posted by Ghayyoor Ahmad in the Vibe Coding is Life group on September 28, 2026. 76 reactions, 299 comments, 12 shares.
The post asks members who do not write code what they do when an AI agent loses track of a long chat and starts adding new bugs instead of fixing the old ones. It drew 299 comments. The most liked answers say to stop relying on long chats: keep sessions short, move project knowledge into instructions, skills and documentation files, and keep the code in GitHub so nothing depends on the chat's memory. One member disagreed and said newer models handle long conversations well, at the cost of more tokens.
The post
Vibe Coders (non-coders): In long chats, AI loses context & adds new bugs instead of fixing them. HOW do you fix this & get the AI back on track?
Why it landed
- It names a problem non-coders run into once a project outgrows a single chat.
- The question is short and specific: one symptom, one ask.
- It addresses non-coders in its first words, which sets who the answers are for.
- It asks for a fix, so the replies came back as steps and habits members could compare.
The best answers from the thread
“You stated the problem in very beginning, stop using long chats. Start using agents and skills and instructions.”
“Use GitHub so the context is externalized and fully commented. Can’t forget when it’s all there.”
“Documentation, and start the chat over when what it already worked on is no longer relevant”
“In my experience, this was more of a problem with the older models, not the new ones: - Opus 5.5 - GPT-6 - Soul - Fable 5.1 - GPT-6 Astra I never worry about previous chats or long context windows. You are probably burning a lot more tokens, but who wants to start a new chat all the time, especially for very complex coding? Just not worth it”
“Anytime I speak to AI, notion starts a meeting - gives me the clif notes - then I follow up on bugs. But I also have local llm agents designed to bug hunt software while building. We do 6 hour reviews every day, update and track log activity, then do a monthly overall review.”
What to take from it
- Start a fresh chat when the earlier work is no longer relevant instead of pushing one long conversation.
- Keep project rules and decisions in instruction and documentation files the agent reads at the start of each session.
- Before a chat fills up, ask the agent to write a hand-off note in a markdown file so the next chat can take over.
- Keep the project in GitHub with clear comments so the context lives in the code, not in the chat.
- If you stay in one long chat on a newer model, expect to use more tokens, as one member noted.
Join the conversation
Post yours. The best ones land in this archive.
This post was shared publicly in the group and is credited to its author, with a link to the original. Want it corrected or taken down? Message the group admins on Facebook and it comes down.