Reference · How-to · ~6 min
How to manage chat memory
Last updated
Keep long chats useful without blowing the context window — in about 6 minutes.
Keep long chats useful without blowing the context window — in about 6 minutes.
Steps
1. **Track usage** — Estimate tokens for system prompt + history + new user message.
2. **Trim oldest turns** — Drop early chit-chat before dropping recent facts.
3. **Summarize mid-chat** — Replace turns 1–N with a short bullet summary.
4. **Store facts externally** — User prefs in a DB; inject only what matters this session.
5. **Re-test** — Ask a question that depends on an old turn; confirm the bot still knows.
Copy-paste prompt
Summarize our conversation so far in ≤8 bullets.
Keep: names, decisions, open questions, policy facts.
Drop: greetings and repeated text.
Then continue answering my next message using that summary as memory.Watch out
Summaries can **drop nuance**. Re-confirm numbers, dates, and names before irreversible actions.