Long AI conversation? Get a summary note, then start fresh.
What the Reel says
Long AI conversation? Get a summary note, then start fresh.
The longer it gets, the more AI forgets.
I changed my team lunch plan midway.
The old plan stayed in the chat.
Prompt: keep only each decision's final version.
The nut allergy made it in.
No old restaurant. Nothing to mix up.
Check the note, paste into a new chat.
Sources and assumptions
- Anthropic: Effective context engineering for AI agents (Sep 29, 2025): 'as the number of tokens in the context window increases, the model's ability to accurately recall information from that context decreases'; needle-in-a-haystack studies 'uncovered the concept of context rot' (checked 2026-11-06)
- Claude Help Center: How do usage and length limits work? (article 11647753, page said 'Updated this week' on 2026-11-06): 'When your conversation approaches the context window limit, Claude summarizes earlier messages to continue the conversation seamlessly.' 'Longer conversations that trigger automatic context management consume more of your usage limit.' 'Try starting a new conversation if you're approaching your usage limit in a longer chat.' Automatic context management needs code execution turned on. (checked 2026-11-06)
- Our test: GPT-5.5 (OpenAI API), Nov 2026 (record.tests ids 'handover-note' and 'new-chat'): The real handover note and the real booking email written from it (checked 2026-11-06)
Assumptions:
- The sample chat is fictional and short (22 lines) and stands in for a long one. The test shows what the handover prompt produces, not that a long chat actually forgot anything.
- The tests ran on GPT-5.5 through the OpenAI API, not the ChatGPT or Claude apps. The Claude facts come from Anthropic's documentation, not from our runs. Label outputs 'GPT-5.5 (OpenAI API), Nov 2026'.
- The handover prompt is general and works in any chat tool. Only the automatic summarizing and usage claims are specific to Claude.
- tests.json also holds two failed Claude entries (ids 'before' and 'after', output null) from the rejected script idea. They are not part of this post; use only 'handover-note' and the last 'new-chat' entry.
- Limitation found by reading the output: the note itself was faithful, but the new chat widened 'a private room' to 'a private room or semi-private area', which nobody asked for. Read the note before you paste it, and the result before you send it.
- Privacy: remove names and personal details before pasting a real chat or document, and check your employer's approved tools.


