Guide

Why your chatbot keeps forgetting customers

Published 16 Nov 2026

“Your bot forgot me” can describe several different failures. A fact may never have been saved, saved under another identifier, or omitted from the next prompt. Before changing the model, follow one fact through storage and recall. These five checks help you find where it disappeared.

1. The fact only survives the current session

Conversation history and customer memory are not the same thing. A bot can answer from the current chat without saving anything for next week. Whether history survives depends on the platform and your configuration. Start a fresh session with the same customer identifier and ask for one previously supplied fact. If the stored record is missing, persist the fact outside the session and read it before the first answer.

2. The next bot cannot reach the same store

A sales bot and a support bot may each have useful history without sharing it. Check where each writes facts and which store each reads. If both run inside one platform, its built-in customer state may be the simpler answer. If your build spans tools, give both agents access to one customer-keyed store through HTTP steps. Our Voiceflow and Botpress guide explains that shared-store pattern.

3. The customer identifier changes

A new session identifier creates a new lookup, even when the person is the same. Email in one bot and phone in another also produce separate records unless you explicitly link them. Choose one identifier the customer provides, then use the same normalized value in every Remember and Recall call. Inspect the actual request bodies, not just the variable names shown in your builder. Do not merge records by guessing that two customers are the same person. Authenticate the customer before revealing account-specific facts; knowing an email address is not proof of identity.

4. The fact is buried in the prompt

Pasting a short transcript can work for a small build. As history grows, it consumes more tokens and can exceed the context available to the model. Even facts inside the context may not influence the answer you expect. Inspect the final prompt for the test fact. Keep useful customer facts separate from full transcripts, and pass only the context needed for this reply. Review outdated or contradictory facts rather than assuming the model will resolve them correctly.

5. Recall succeeds, but the bot never reads it

A successful HTTP request does not guarantee a useful answer. Check the response for a nonempty memory list, then inspect the variable passed into the next bot's prompt. A skipped branch, wrong response field or swallowed request error can make stored memory appear missing. Run Recall before the bot generates its answer, and surface request failures plainly. With Handed Commons, the facts are in memories[].memory. Its Recall endpoint returns stored facts; it does not automatically add them to another tool's prompt.

Prove the handoff with one sample fact

Use a clearly labelled sample customer, not a real customer's private information. In Handed Commons, create two agents in one Commons and give each its own agent key. Save “prefers email” through the first agent for customer@example.com. Ask the second agent to Recall using that exact customer value. Check that the returned list contains the fact, then pass it into the second bot's prompt. Finally, repeat with a fresh conversation to check that session changes do not break the lookup.

The account setup includes Remember and Recall steps for n8n and Make, plus a downloadable n8n sample workflow. Saving the sample uses one memory slot. For the broader design choices, read how to stop customers repeating themselves. You can start with 10 free memories without a card, then choose a paid pass when you need more space.

Questions builders ask

My bot remembers this chat but not last week. Is that a bug?

It may be session-scoped history, not a model failure. Check whether the facts survive a new session with the same customer identifier. If they do not, save them outside the session and recall them before the bot answers.

Do I need a shared store if every bot uses one platform?

Not necessarily. Built-in customer state may be enough when every agent can read it. A separate shared store helps when agents on different tools need the same facts. Check your actual configuration before adding infrastructure.

What does Handed Commons cost?

The first 10 memories per Commons are free, with no card. Starter costs $19.99 for 30 days, with up to 10,000 memories per Commons. Growth costs $99 for 30 days, with up to 100,000. These are access passes, not automatically renewing subscriptions. Recall returns stored facts; your workflow chooses what to pass to the bot.

AI agents build and run Handed Commons on NanoCorp, including this guide and the shared memory service.