Maaz Zindani

Loading the index…

↑↓ move↵ openesc close4072 pages

Help for owners

My AI agent forgets what the customer said before

Short answer

A model has no built-in memory between requests. It only sees what your system sends: the system prompt, the earlier messages and tool results, within a limited context window. Remembering a customer across chats means storing the facts yourself and sending the right ones back in.

  • Written for business owners
  • Updated
  • By Maaz Zindani, Senior Software Engineer

Does this sound like your situation?

  • The agent asks the same questions again.
  • It does not recall yesterday's chat.
  • In long conversations it forgets the start.
  • It mixes up two customers.
  • Answers get worse as the chat gets longer.

Why does this happen?

  1. The context window is its working memory

    Anthropic describes the context window as all the text the model can reference when responding, a working memory. Everything in the request counts towards it, including system prompt, messages and tool results.

  2. Each turn carries the earlier turns

    History is accumulated by resending previous turns. If your system does not send them, the agent starts fresh.

  3. Long chats lose accuracy

    Anthropic notes that as token count grows, accuracy and recall degrade, so more context is not always better; curating it matters.

  4. Nothing is saved between sessions

    A new chat is a new request. Facts such as preferences and open orders must be stored in your own database and loaded.

  5. Customer records are mixed up

    If a lookup returns the wrong profile, the agent confidently uses the wrong history. Matching identity correctly is a data task.

What can I do about it today?

  1. Write down what must be rememberedName, open orders, preferences and past decisions, not everything.
  2. Decide how long to keep itCheck your privacy duties before storing conversations.
  3. Check where chats are stored todayAsk whoever built it whether history is saved and where.
  4. Do not paste private data into free toolsKeep customer chats inside tools you have agreed terms with.
  5. Test with a repeat customerRun a second conversation and see what is missing.

When should I bring someone in?

  • Customer records are involved.
  • You must follow rules on storing personal data.
  • The agent works across email, chat and phone.

How do I handle it?

  1. Separate memory typesShort-term chat, saved customer facts and company documents are handled differently.
  2. Store facts in your databaseA short, structured record per customer, updated by the agent and visible to you.
  3. Trim and summarise long chatsOlder turns are condensed so the context stays focused.
  4. Identify the customer correctlyThe same person on different channels maps to one record.
  5. Let customers and staff correct memoryA wrong saved fact can be edited or removed.

The price depends on what I find, so I quote in writing after a short discovery call. Tell me what is happening and I reply within 24 hours.

What do owners ask about this?

Can I just buy a bigger context window?

A bigger window raises capacity but Anthropic notes accuracy can still degrade as it fills, so choosing what to include still matters.

Is saving chats allowed?

It depends on your country and what you told customers, so ask your adviser.

Sources

How was this page made?

I drafted this page with AI assistance from the official documents listed under Sources. Found a mistake? Tell me.

Next step

Want this off your desk?

Send me a few lines on where things stand. I reply within 24 hours.