Which AI Support Tools Actually Remember Customers Between Visits

Search whether an AI support agent remembers a returning customer and you get nine results explaining how to build memory yourself, written by companies selling memory infrastructure and by agencies selling the build. Not one of them says which support platform already does it.

So we read the documentation. The answers are published, they are wildly different, and almost nobody puts them next to each other.

The distinction vendors blur on purpose

There are two different things called memory. Context within one conversation, so the agent does not forget what you said four messages ago. And context between conversations, so a customer returning on Wednesday does not re-explain Monday's problem.

The first is table stakes and every product has it. The second is rare. Marketing copy tends to say "remembers the context of the conversation", which is true of the first and reads like a promise about the second.

What each platform publishes

Read on each vendor's own documentation, August 2026.

Fin is the only one that has actually built it

Intercom's Fin publishes the mechanism in detail. Its documentation says Fin Memory lets Fin reference what it already knows rather than starting from scratch every time, and it names three components with separate retention: recent messages and episodic summaries kept for 45 days, and facts a customer explicitly shared kept for 180.

It is on by default for every Fin for Service workspace. It is also, in Intercom's own words, currently in beta, limited to identified users, and applies to messenger and email channels only. Notably it does not work for anonymous visitors, and neither teammates nor customers can see or edit what it holds.

That last constraint is the one to take to your legal team. Intercom flags it directly: because the feature retains and reuses personal details customers share, you may need to update your own privacy notice.

And Fin's own FAQ contradicts the feature page

On the same site, the Fin FAQ says that while Fin learns from successful interactions over time, it does not retain memory of answers within a single conversation. Intercom's parallel copy of that FAQ says the same thing in slightly different words.

Neither mentions Fin Memory. Searching Intercom's own help centre for the term returns nothing, while the Fin site carries the memory article as a top-level entry. One product, two domains, and the older domain does not document the newer feature.

Three hours, and the honesty of saying so

Tidio publishes the tightest window in the category and states it plainly: Lyro can see three hours into the past for the most recent history, covering conversations with live agents as well as flow and AI interactions.

Three hours is short. It is also the most useful number any vendor here publishes, because it answers the question without needing interpretation. A customer returning the next morning is a stranger, and Tidio does not pretend otherwise.

Zendesk documents the opposite of memory, precisely

Zendesk's documentation explains what happens when a customer comes back after a gap. Bots treat a conversation as expired after 72 hours of inactivity and delete the metadata. When the customer returns, the system cannot find the identifier it needs, so it grabs the last ten messages from previous interactions and appends them to the ticket to provide some context.

That is a fixed-size fallback, not memory. Ten messages, whatever they happened to be, stapled to a new ticket.

Zendesk does document persistent conversations — but read who the subject is. The benefit described is that agents can refer to earlier interactions, in Zendesk's sense of human agents. It is a feature for the person, not the software.

Ada is the only vendor that documents when the slate is wiped

Ada's documentation is the most detailed in the set and what it documents is the absence. Global variables persist throughout an interaction and reset when a new conversation begins. Conversations end automatically after about 24 hours on chat and 72 on email, and after that the next message creates a new conversation entirely.

What does persist is an identifier rather than content, and Ada is candid that even it is fragile: clearing cookies produces a new identity. On the returning-customer case its documentation says persisted variables may still be available — the only hedge of its kind we found in any vendor's docs, and a more honest one than a confident claim.

The pattern worth the whole piece

Three platforms document no cross-session memory at all: a search of the Gorgias AI Agent documentation, the Chatbase documentation index and the Freshworks AI Agent Studio docs returns no mention of memory in any of them.

Gorgias has something adjacent — a Customer Timeline showing past tickets and order data. Every verb in that documentation has a human subject: review, catch up, respond. And the same page notes the identity problem directly, that a customer might arrive from a channel with no email address, like an Instagram DM, in which case the tickets do not land under one timeline at all.

Freshworks is the clearest case of a word doing double duty. Its documentation defines a session as 24 hours on chat and 72 on email — and defines it there because a session is the billing unit. The window is published because they sell against it, not because it describes what the agent recalls.

The two loudest claims are the two you cannot check

Sierra and Decagon make the largest memory claims in this category. Sierra's blog says its data platform unifies everything your company knows about a customer across sessions, channels and systems, and calls it the memory and intelligence layer. Decagon's says a customer can step away and come back days later without re-explaining a thing.

Neither publishes documentation a buyer can check. Both doc sites are behind a login. No retention period, no statement of what is stored versus discarded, no statement of whether identity is automatic.

Every specific, falsifiable number in this piece — Fin's 45 and 180 days, Zendesk's 72 hours and ten messages, Ada's reset, Tidio's three hours, Freshworks' 24 and 72 — came from a vendor that publishes its documentation openly. The correlation runs one way and it is not an accident.

What to ask

Does the agent read anything from a conversation that ended last week, and if so what and for how long? Does it work for a customer who is not logged in? Is memory on by default or opt-in, and can we turn it off? And where is that written down.

A vendor that answers the last question with a link is telling you more than one that answers the first three enthusiastically.

Read from each vendor's own documentation in August 2026. Where a vendor's marketing and documentation disagree, both are quoted. Where documentation is behind a login we say so rather than treating a blog post as a specification.

WRITTEN BY AR · UPDATED 2026-08-20

I read the fine print. Vendor pricing pages, billing definitions, terms, funding filings and acquisition notices — then I do the arithmetic nobody publishes: what a platform actually costs at your volume, what its headline metric is really counting, and who owns it now. I do not run benchmarks, and no page here pretends otherwise.

Editorial policy