Start fresh when a conversation gets long
Long conversations make agents forget early details. Learn the signs, and how to move to a new conversation without losing what you've built up.
An agent can only hold so much of a conversation in mind at once. That space, called the context window, fills with every message, every file the agent reads, and every command it runs. When it’s nearly full, the agent summarizes older parts to make room. Claude Code’s documentation explains the catch: “instructions from early in the conversation can get lost.”
Signs it’s time
- The status says Compacting. Claude Code, Codex, and Grok Build conversations show Compacting under the title while the agent summarizes. In the agent activity panel, this appears as Compacting conversation, then Conversation compacted. One compaction is normal. Several in one conversation mean it has grown very long.
- The agent forgets things you settled earlier, like a color you chose or a page it shouldn’t touch.
- You’ve corrected the same thing more than twice. Claude Code’s guide says the conversation is then “cluttered with failed approaches,” and that “a clean session with a better prompt almost always outperforms a long session with accumulated corrections.”
- The next task is unrelated. A new task deserves a new conversation. See Use one conversation per task.
Don’t start over for follow-ups on the same problem. OpenAI’s Codex guide notes that staying in the same chat “preserves the reasoning trail.”
How to carry context over
- Ask for a handoff doc. In the old conversation, ask the agent to save a summary as a doc. Agents can create docs in your project with Enjoy’s tools, and the doc appears in Files.
- Mark the old conversation completed. See Mark finished work as completed.
- Start a new conversation and mention the doc. Select New conversation, type @, and choose the doc under Docs. Enjoy puts a link to the doc in your message. Ask the agent to read it, as in the example below. Agents open docs with Enjoy’s document tools.
You can also ask the new agent to read the old conversation by its name. Agents can list and read the project’s conversations, but a short doc is faster to read and easier for you to check.
For rules that apply to every task, such as “always use our brand colors,” put them in your project’s Instructions instead. Enjoy sends them to the agent in every conversation, so you don’t have to repeat them. See Write instructions your agent will follow.
An example
In a long conversation about a booking page, the status keeps showing Compacting, and the agent has undone a layout you’d approved. You send:
“Before we stop, write a doc called ‘Booking page handoff’. Include what works now, what’s left to do, decisions we made and why, and approaches that didn’t work. Keep it short.”
Then you mark the conversation completed, start a new one, and send:
“Read @Booking page handoff and pick up from there. Next, add a confirmation email after someone books.”
Common pitfalls
- Starting over without a handoff. The new agent has to rediscover everything, which costs time and usage.
- Pasting the whole old conversation. That brings the clutter with it. A summary of decisions and open work is enough.
- Keeping the handoff doc forever. Once the work is done, archive it so it doesn’t mislead later conversations. See Keep project knowledge in docs.