Caller Memory

Caller Memory lets an agent remember facts about a specific caller -- a preference, an ongoing issue, anything worth recalling -- and bring them up naturally the next time that same phone number calls this agent again. It replaces the manual workaround of wiring a Pre-Call Data Fetch URL to an external CRM just to get basic call-to-call continuity.

Facts are scoped per agent, not shared across your organization's other agents. If a caller reaches two different agents (e.g. a billing line and a scheduling line), each agent only ever recalls what it itself learned -- not what the other agent knows about that same caller.

Enabling it

Open an agent's Settings and toggle Caller Memory on. It's off by default for every agent -- nothing is stored or recalled until you enable it.

How it works

Asking permission. Before the agent remembers anything about a caller it hasn't asked before, it asks for their permission in the conversation itself -- something like "is it okay if I remember that for next time you call?" -- and silently records the answer. This isn't just a prompt instruction: Talkr enforces it in code, so a fact is never stored unless that caller's permission is on file, independent of whether the model happens to ask first.

Remembering. Once permission is granted (this call or a previous one), an instruction in the agent's system prompt asks the model to silently call a remember_caller_fact tool, in parallel with its normal spoken reply, whenever the caller shares something worth recalling on a future call -- not routine details only relevant to the current one. This is the same silent-tool-call mechanism Emotion Detection uses: it never adds a round trip, and the caller never hears or sees any indication it's happening.

Recalling. At the start of every call, before the agent's first reply, Talkr checks whether this caller has granted permission. If they have, it looks up any facts previously remembered about their phone number and makes them available as the {{caller_memory}} template variable -- see Context & Variables. Reference it directly in a node prompt (the Global node is a natural place) so the agent can use it naturally:

Known history with this caller: {{caller_memory}}

On a caller's first call, when permission hasn't been granted, or when Caller Memory is off, {{caller_memory}} is simply empty.

This only applies to voice calls with a real caller phone number -- text chat and other callerless contexts have nothing to match facts against, so the tool and the lookup are both silently inactive there.

Declining and revoking

If a caller declines when asked, the agent won't ask again on that same call -- it only revisits the question if the caller brings it up themselves, on that call or a later one.

A caller can also revoke permission at any point by asking the agent to stop remembering them or to forget what it knows -- for example, "please forget everything about me." The agent records this immediately and deletes every fact stored about them on this agent, the same way the admin-side "Forget a caller" action does (see below).

Viewing and managing consent + remembered facts

Every agent with Caller Memory has its own management page (linked from the Settings toggle) listing every fact it has remembered, grouped by caller, along with each caller's current consent status. From there you can:

• Delete a single fact
• Forget a caller entirely -- deletes every fact stored for one phone number on this agent and revokes their consent, so the agent asks again before remembering anything new about them

Deleting a single fact only removes that fact; it doesn't change the caller's consent status or affect calls already completed.