Support
Memory for customer support AI
The customer explained it in March, again in June, and is explaining it now. Recall what this account has already been told, with the ticket it came from.
Today
The assistant sees one ticket. The customer's history is in the helpdesk, the shared mailbox, and a call somebody recorded, and none of it reaches the model. So the customer repeats themselves, gets told something that contradicts what they were promised in a different channel, and the promise that was made is discoverable only by whoever remembers making it.
With a memory
Everything this account has been told is one query away, with the channel and date it was said in. The assistant answering today can see the commitment made in a call in March, and can tell the customer something consistent with it. When two channels disagree, that shows up as a conflict rather than the assistant confidently repeating whichever it happened to retrieve.
What that looks like
Promises are findable
A commitment made on a call is a memory with the recording as its source, so nobody has to remember that it was made in order to honour it.
Escalation carries context
The human taking over gets what the assistant knew, in the same form the assistant had it, instead of a transcript to read from the top.
Access rules survive the recall
A memory captured from a private internal thread is not returned to a customer-facing surface. The authorisation from the source is carried through rather than reapplied by hand.
The shape of it
const history = await client.search.query({
query: "billing arrangement",
spaceIds: ["spc_northwind"],
limit: 8
});
// Worth reading. Search DEGRADES rather than fails: with embeddings
// down it falls back to deterministic retrieval and still answers.
if (history.diagnostics.degraded) log.warn("recall was narrower than usual");
// Each result carries where it came from and when, so a reply can say
// "as we agreed on the 12th" and be checkable.
for (const { memory } of history.results) {
console.log(memory.content, memory.sources, memory.temporal);
}What this does not do
It will not stop an assistant from being wrong about your product. Memory makes it consistent with what was actually said before, which is a different property from being right, and both are worth having for different reasons.