Chatbase Remembers the Conversation. It Forgets the Customer.

You train a Chatbase agent on your help docs, wire it to Stripe and Zendesk, and drop it on your site. It answers questions well. Customers get order statuses, book calls, open tickets, all without a human in the loop. Then one of them comes back a week later and asks, "Hey, what happened with my refund?" The agent has no idea who they are, what they asked about last time, or that the refund was ever discussed. The chat log is sitting in your dashboard. The agent itself is starting from zero.

That is the shape of Chatbase memory: excellent within a conversation, absent between them.

What Chatbase actually remembers

Chatbase agents keep full multi-turn context inside a single conversation. If a customer mentions their order number in message three, the agent can still use it in message thirty. Follow-ups work, clarifications work, and the agent does not ask the same question twice while the chat is open. For a support widget, this covers most of what a single interaction needs.

The agent also sits on top of your knowledge base. Upload PDFs, docs, and URLs, and it retrieves relevant answers from them with source grounding. That is retrieval, though, not memory: it tells the agent what your business knows, not what this customer did.

On top of that, newer Chatbase agents can take actions through integrations: check an order in Stripe, create a ticket in Zendesk, book through Calendly, search the web. The agent can do things. But doing things is not the same as remembering that it did them. An action leaves a record in your systems; it does not leave a memory the agent can recall next week.

What it does not remember

Across conversations, Chatbase starts blank. A returning customer is, for all practical purposes, a stranger. There is no persistent profile the agent builds up over time: no record of preferences stated last month, no memory of the bug report from Tuesday, no sense that this person has talked to you five times before.

Your dashboard does keep conversation history. You can open last week's chat and read every word. That is a log, not a memory. Nothing in that log is available to the agent the next time the customer writes in, unless a human reads it and re-states it. The knowledge is in your building, but the agent cannot walk over and look at it.

This distinction matters more than it sounds. Session memory makes one conversation feel smart. The absence of cross-conversation memory makes the relationship feel dumb. Customers notice the second one more than the first. Nobody praises an agent for remembering what they said two minutes ago; everybody notices when it forgot what they said two days ago.

Where this bites hardest

The damage shows up in three places. First, repeat customers. Someone with an ongoing issue, a subscription question, or a complex setup talks to you more than once. Every return visit restarts the intake: name, account, what is broken, what was already tried. Each restart costs the customer patience and costs you the resolution you were close to last time.

Second, follow-ups. "I asked about this last week" is one of the most common sentences in support. When the agent cannot connect last week to this week, the customer either re-explains everything or gives up and demands a human. Both outcomes erase the automation savings the agent was supposed to deliver.

Third, personalization that never compounds. An agent that remembered a customer's plan tier, their integration stack, and the workaround that fixed their issue in March would answer like someone who knows them. Without that, every answer is generic-first and specific-only-after-interrogation. The agent works, but it never gets better at its job.

Four ways to close the gap

Option 1: Make the customer do the remembering. Ask returning visitors to restate their context every time: account email, order number, what happened before. This works in the sense that the agent gets the facts. It fails in the sense that customers hate it. You are taxing the exact people who came back.

Option 2: Key conversations to identity in your own database. If your widget passes a user ID or email into the conversation, you can store summaries of past chats in your CRM or database and inject a short brief into each new conversation: who this is, what happened before, what is unresolved. This is the most common DIY fix, and it genuinely works. It also means you are now building and maintaining a memory system: writing the summarizer, storing the records, keeping them fresh, handling deletions when someone asks to be forgotten. The engineering is not hard; the ongoing maintenance is.

Option 3: Let the handoff carry the context. Chatbase escalates to a human with the conversation attached, so at least the human agent does not start blind. This protects the human-handled cases. It does nothing for the fully automated ones, which are the ones you bought the agent for.

Option 4: Give the agent a real memory layer. Instead of building your own store, connect the agent to shared memory that persists across conversations and tools. Vilix AI is built for exactly this: it is cloud-hosted, so there is no infrastructure to run, and it connects over MCP, so the same memory follows the agent everywhere it runs. It stores full conversation history, not just extracted facts, which means the agent recalls what actually happened rather than a lossy summary. When a returning customer asks about their refund, the agent can pull up the thread instead of asking them to start over.

Vilix AI is free forever on the free plan, with a 7-day Pro trial that needs no credit card. If you want to try the approach, you can start here: https://vilix.ai/?utm_source=vilix-blog&utm_medium=article&utm_campaign=chatbase-remembers-chat-forgets-customer. And the exit is clean: export everything in a portable format or delete it all instantly, any time. Memory you cannot leave is not really yours.

The bottom line

Chatbase does what it promises inside a single conversation: it follows the thread, grounds answers in your docs, and can take real actions. What it does not do is build a relationship with the person on the other side. Every conversation is a first conversation.

That is fine if your support is mostly one-and-done questions. It becomes expensive the moment customers come back, which, for most businesses, is the whole point of having customers. The fix is not a better chatbot. It is memory that outlives the chat window.