Knowledge base & memory

Your agent remembers the last call.

Most voice AI starts every call from zero — so your returning customer gets interrogated about the unit you fixed for her in the spring. Voicent builds a knowledge base out of the calls, records, and job history you already have, then hands the agent the few facts that matter for the call it's on right now.

How it works

Any source. One graph. The context for this call.

Every call, CRM record, and closed job becomes memory. Voicent turns them into a graph of who the caller is, what she owns, what you did for her, and what changed. Before the agent speaks, it pulls the handful of facts that bear on this call — not a transcript archive, and not a prompt stuffed with everything you know about her.

Sources you already have

  • CRM

    Your CRM records

    Contacts, addresses, opportunities, and the notes your team already keeps.

    GoHighLevelHubSpotAttio

  • JOBS

    Job & invoice history

    What you fixed, when, and for how much — from the system that already tracks it.

    ServiceTitanHousecall ProQuickBooks Online

  • DOCS

    Plans, policies & price book

    The rules the agent has to know: what a plan covers, what a job costs.

    NotionGoogle SheetsAirtable

  • CALLS

    Every call you take

    What was asked, what was promised, what was booked — written back when the call ends.

Voicent builds the graph

ownsservicedcovered byprefersCALLERPriya RamanASSETRooftop AC unitPLANMaintenance planJOBCapacitor replacedPREFERENCEMorning windows

Who she is, what she owns, and what changed.

The context for this call

context_block.txtassembled before the agent speaks

# FACTS

·Priya Raman owns the rooftop unit at 114 Alder Ct.valid 2023-06-02 → present

·Its capacitor was replaced last spring.valid 2026-04-18 → present

·She renewed the annual maintenance plan.valid 2026-03-01 → 2027-03-01

×Out of contract — pay per visit.invalidated 2026-03-01, superseded by the renewal

·She books morning windows — two afternoon slots were missed.learned on the last call, 2026-05-02

# THE AGENT OPENS WITH

“Welcome back, Priya — is this about the same unit we serviced in the spring?”

Priya is an example caller. The structure is not: this is the shape of what a Voicent agent reads before it opens its mouth.

Memory that keeps up

A fact isn't true forever. It's true for a while.

Facts get an end date, not a delete

When Priya's plan renewed, the old fact didn't vanish — it was marked invalid from the day it stopped being true. The agent can't quote a fact that expired, and you can still see that it was once true.

Retrieval, not recitation

The agent never gets the whole archive. It gets the handful of facts that bear on this caller and this call, which is why a returning caller doesn't add thirty seconds of throat-clearing to the front of the conversation.

Written back when the call ends

The last step of the workflow saves what the call learned. That is why the fourth source on this page is the calls themselves — memory that only reads is just a database.

Memory is a step, not a black box

Look at any workflow on this site and you'll find memory where you'd expect it: a lookup before the agent greets anyone, and a save after the call ends. It runs on the Zep Memory connector — 7 actions your workflows can call, like any other integration in the catalog.

Browse all workflowsThe Zep Memory integration

Access, retention & audit

Memory you can govern.

An agent that remembers your customers is only an asset if you control what it keeps, how long it keeps it, and whether you can explain what it said. Those aren't settings buried behind a support ticket — they're decisions we make with you while the workflow is being built.

ACCESSScoped to your workspace — and to the workflows you allowLookups are keyed by caller and by workspace: one business's graph is never another's context. And nothing is remembered by accident. If a workflow has no memory node, that call writes nothing.
RETENTIONYou decide how long a caller stays rememberedHow long call memory lives, what it keeps, and what gets dropped when a customer asks to be forgotten are settings we configure with you when the workflow is built — not defaults you inherit and discover later.
AUDITEvery fact traces back to the call that produced itEach fact carries where it came from and the window it was true for, and memory is read and written at steps you can see in the diagram. When someone asks why the agent said what it said, the answer is in the graph — not buried in a prompt.

Questions

Does the agent actually remember callers between calls?

Yes. The workflow opens by looking the caller up by number and closes by saving what it learned, so the second call starts where the first one ended — by name, with the last job in hand.

What happens when a fact changes?

The old fact is invalidated with a date rather than overwritten. The agent answers from what's true today, and the history of what used to be true stays intact.

Does it read my entire CRM on every call?

No. It retrieves the facts relevant to the caller in front of it. That keeps the call fast and keeps the agent from reciting things the caller never asked about.

Can I see what the agent knows?

Memory reads and writes are ordinary workflow steps, so you can see exactly where the agent touches memory on any template diagram — including the ones on this site.

Can you delete what the agent remembers about someone?

Yes. Memory is stored per caller, so a caller's facts can be removed on request. How long memory is kept in the first place is a retention setting we agree with you when the workflow is built.

Give your callers an agent that already knows them.

Book a consultation and we'll map what your agent should remember — and show you the workflow that reads it, uses it, and writes the next call back.