Does Decagon Remember Between Conversations? User Memory, Honestly Explained

Yes. Decagon added a dedicated User Memory feature in spring 2026, and it is the real answer to the memory question: conversational memory carries unstructured chat history across sessions so returning customers pick up where they left off, and the platform extracts signals like feature requests, sentiment, and stated preferences into structured metadata on the customer profile. Memory calls are wired into Agent Operating Procedures, with redaction, expiration, and audit controls on top.

That is a bigger deal than it sounds, because the answer used to be no. Before this launch, Decagon's agent was stateless between conversations the way most platform agents are: brilliant inside a chat, a stranger when the customer came back. The Spring 2026 update changed the architecture, and the details matter more than the announcement.

What does Decagon's user memory actually store?

Two different things, and the distinction is the whole product.

First, conversational memory: the unstructured history of what the customer actually said. Decagon describes the user-level abstraction as memory that builds across interactions. In practice, a customer can start troubleshooting an issue, step away because life happens, and come back days later without re-explaining anything. The agent still has the context of the unresolved issue, which is what makes the follow-up conversation short instead of painful.

Second, personal context: the platform extracts key user signals and stores them as structured metadata attached to the customer's profile. Feature requests they mentioned. Sentiment from past interactions. Stated preferences, the shirt size in the shopping example Decagon itself uses. This is not transcript search; it is a distilled view of who the customer is, kept current.

Then there is the user view Decagon gives your team: a single profile per customer that pulls together conversation history, outreach attempts, and stored preferences, with tags from quality-monitoring flags, CSAT scores, and whatever custom signals you configured. Instead of digging through transcripts to understand who you are dealing with, the relevant context sits at the top of the profile.

How does the memory get used inside a conversation?

Through Agent Operating Procedures, Decagon's AOPs. These are the playbooks that define how the agent behaves, and the important mechanical detail is that you specify memory calls directly inside each AOP. Decagon calls these context-aware AOPs: stored memory and user data are taken into account in real time, so the interaction personalizes itself from the customer's memory as it runs.

That is the right place to put it. Memory that sits in a database nobody queries is decoration. Memory wired into the procedure layer is behavior.

The governance story is also unusually complete for the category. You decide what gets stored and for how long, with built-in redaction and expiration policies. Memory storage is opt-in, customers stay inside your existing privacy and consent flows, and audit logs record what is stored for compliance. Data is fully portable both ways: you can import static customer data in and export interaction data out, so the customer snapshot stays in sync with the rest of your stack instead of becoming a second system of record you cannot leave.

Where does Decagon's memory stop?

Here is the honest boundary, and it is the part most teams miss when they evaluate platforms on memory alone.

Decagon's memory remembers your customers for the agent Decagon runs. It is scoped to the customer relationship inside Decagon's platform: chat, email, and voice channels that Decagon operates. That is exactly what a support team needs, and it is what Chime's COO was describing when she pointed to cross-channel memory as the reason every interaction stays connected.

But there are three things it does not do, and all three are where automation operators actually get bitten.

First, it does not remember across tools. If your support flow also runs through an n8n workflow, a scheduled reconciliation agent, or a human rep working in a different system, Decagon's user memory does not follow. It is the platform's memory, attached to the platform's agent.

Second, it does not remember your operation. Decagon's memory is about the customer, not about you. The failed run at 3 AM, the escalation that exposed a gap in the AOP, the decision your team made about how to handle billing disputes: none of that lives in user memory, because that was never its job.

Third, it does not help the agents you build yourself. If you ship your own agents on your own stack, alongside the Decagon deployment or independent of it, you start from zero on memory again.

What about the agents you build yourself?

That third gap is where a hosted memory layer earns its keep. When your own agents need the same kind of persistence, the options are the usual three: bolt a database onto every agent and maintain the retrieval plumbing yourself, depend on each platform's built-in memory and accept that every platform starts over, or put one memory layer underneath everything.

Vilix AI is built for that third option. It is cloud-hosted, so there is no infrastructure to manage, and a built agent connects with an API key as a Bearer header to https://api.vilix.ai/mcp and gets full memory with zero memory infrastructure. The same memory is available everywhere over MCP, which means the agent you deploy today and the scheduled flow you add next month share one context instead of two blank slates. It stores full conversation history, not just extracted facts, and your data stays portable: export everything or delete it anytime. It starts with a free plan that is free forever, and a 7-day Pro trial with no credit card.

Decagon made the right call for its agents: memory is infrastructure, not a feature bolted on later. The teams running their own agents should make the same call, and they should make it once instead of reimplementing it inside every platform.

Frequently asked questions

Did Decagon always remember customers between conversations?

No. The dedicated User Memory feature arrived with the spring 2026 update. Before that, the agent was context-aware inside a conversation but did not carry a persistent user profile across them.

Does Decagon's memory work across chat, email, and voice?

Yes. User memory is built on a user-level abstraction that spans channels and sessions, which is what lets a customer start on chat and continue by phone without re-explaining.

Who controls what Decagon stores about a customer?

You do. Storage is opt-in, you set redaction and expiration policies, and audit logs track stored conversations. Customer data stays inside your existing privacy and consent flows.

Can you get your data out of Decagon's memory?

Yes. Decagon describes full data portability: import static customer data from your systems, export interaction data back out, so the memory does not become technical debt you cannot leave.

Does Decagon's memory help agents I build on my own stack?

No. It is scoped to the agents Decagon runs inside its platform. Your own agents need their own memory layer underneath them.

If your agents wake up blind every run, give them one memory. Vilix AI is the shared memory layer for AI agents: cloud-hosted, same context everywhere over MCP, full conversation history. Free forever, no credit card.