Does Voiceflow Remember Between Conversations? What Actually Persists

You built a Voiceflow agent. It handles the first conversation fine. A week later the same person comes back and asks about something they told you last time, and the agent answers like it has never met them. The name it remembered yesterday is gone. The preference it confirmed three days ago never happened.

So you ask the obvious question: does Voiceflow remember between conversations?

The short answer is yes, but only the things you told it to remember, in the way you told it to remember them. Nothing is automatic. Understanding the boundary between what persists and what resets is the difference between an agent that feels continuous and one that starts blank every visit.

What persists: user variables

Voiceflow's main mechanism for cross-visit memory is the variable. Every project has built-in variables, and you can define your own. The documentation is explicit about their role: keep them on "for long-lived details you want your agent to remember across visits, like a user's name or their preferences."

Two built-ins do quiet work here. user_id is the stable identity Voiceflow keys state to, set through the chat widget or the API. For phone integrations it is the caller's phone number on a call. sessions counts how many times that user has opened the agent. So Voiceflow knows the user is back, and anything you stored in a variable during their last visit is still there waiting.

This is real persistence. Set a variable during Tuesday's chat and Wednesday's session can read it. That covers the most common "remember me" cases: names, account tiers, the plan they chose, the ticket they were tracking.

What resets: the conversation itself

What does not carry over is the conversation. Voiceflow exposes a built-in variable called vf_memory, and its definition matters: "the last ten user inputs and agent responses as a string, including tool calls." Ten turns of the current conversation. Not the last session. Not a summary of everything the user ever said. Ten turns, right now.

The AI steps in Voiceflow have a Memory option that works the same way. Enabling it includes the previous 10 turns alongside your prompt so the response has context. Turn it off and each AI step reasons in isolation. Either way, the memory window is the same: the recent tail of the active conversation.

Long conversations hit a second ceiling. Voiceflow's own community guidance warns that token limits can drop context: the platform saves the last 10 turns, but when a conversation runs long, earlier context can be lost. A single marathon session can forget its own beginning.

And here is the detail that surprises operators: the Conversations API docs note that variables are not automatically reset when a conversation's state is reset. That sounds convenient until you realize the inverse is also true. Variables keep their last value whether it is still true or not. There is no lifecycle on a remembered fact. Nothing expires a stale preference or flags a contradiction. Your agent remembers everything you wrote and understands nothing about it.

The gap operators actually hit

Put together, Voiceflow's memory model is a whiteboard that survives the session but never gets cleaned, plus a ten-turn notepad that gets shredded when the conversation ends. Every fact worth keeping must be individually engineered into a variable, updated at the right moment, and maintained by your own logic forever.

Three things fall through this gap:

First, facts you did not think to store. A variable only holds what your flow explicitly writes. If the user mentioned a deadline in passing and your flow never captured it, it is gone with the session.

Second, the narrative of what happened. Variables hold values, not stories. "plan_tier = pro" survives; the forty-minute conversation where the user explained why they upgraded does not. The reasoning, the objections, the context that would change how your next agent treats them is lost.

Third, anything outside Voiceflow. The variable lives inside Voiceflow's runtime. The same user's n8n workflow, their scheduled agent, the phone call they made, the email they sent: none of it is in that variable. Each tool keeps its own whiteboard, and the operator keeps the only complete picture, in their head.

This is the pattern behind almost every "my agent has amnesia" complaint. The platform remembers per design, but the memory is scoped, manual, and siloed. What operators actually want is a memory that works like a human colleague's: everything that was said, in every tool, available the next time it matters, without a flow chart deciding what counts.

What a memory layer adds on top

Voiceflow is the conversational front end. Underneath it, a hosted memory layer changes what "remember" means. Instead of hand-picking facts into variables, the agent stores the full conversation and recalls the relevant pieces when a trigger fires: the user's question, a new session starting, a scheduled run waking up.

Vilix AI (vilix.ai) is built for exactly this. It is cloud-hosted, so there is nothing to run or maintain. The same memory follows the user across Voiceflow, n8n, scheduled agents, and every other AI tool through MCP, so context is not trapped in one platform's runtime. It stores full conversation history, not just extracted facts, which means the narrative survives, not only the values. The free plan is free forever, a 7-day Pro trial needs no card, and everything can be exported or deleted at any time.

The practical shape is simple. Voiceflow handles the conversation, the variables still do their job for immediate flow logic, and Vilix AI holds the long-term memory that no single platform's variable system was designed to carry. When the user comes back next week, the agent does not just know their name. It knows the story.

The honest summary

Voiceflow remembers between conversations the way a good form remembers: the fields you filled in are still there, keyed to who you are. It does not remember the conversation itself, anything you never wrote into a variable, or anything that happened anywhere else. If that is enough for your use case, the variable system is doing its job. If your agent needs to know what happened, not just what got recorded, that memory lives in a layer built for it.