Talqing keeps everything forever by default. One workspace setting changes that, and when the deadline passes the content of a call is deleted while the record of it stays — because an invoice has to keep resolving.

​
The retention policy

retention_days is a single setting on the workspace: Settings → General in the dashboard, or PATCH /v1/org. Saying that plainly: unless you set a policy, nothing is ever deleted automatically. No recording ages out, no transcript expires, and no bucket lifecycle rule quietly removes an object behind the database’s back. A recording is gone only because something deleted it and stamped the record. Setting a policy is the deliberate act. See workspaces and roles for who can perform it.

​
What a purge deletes

When a call’s deadline arrives, its content is deleted. This is not an audio feature — a transcript is personal data in exactly the way the audio is, and sharper: it is indexed, machine-readable, and the copy that travels into analysis, webhooks and your language-model provider. So a purge takes all of it. Deleted:
  • The transcript — every message, in both directions.
  • Every tool call the agent made and everything those tools returned.
  • The session’s userdata, its summary, its outcome rationale and its extracted analysis fields.
  • The recording, and the screen recording if the agent watched a screen. Both are content in the same sense, and a screen video surviving a purge would be the most indefensible thing here to get wrong.
  • Every image anyone attached to the conversation.
  • The whole event trace, except one record: if a caller withdrew consent to being recorded, that one event survives, carrying a single timestamp and nothing derived from the conversation. It is the evidence the content was handled correctly, which is precisely what someone asks for after the content is gone.
  • The phone numbers on the call — the caller’s, the agent’s, and any number the call was transferred to.
Survives:
  • The call record itself: its duration, status, outcome, close reason and every cost column.
  • The usage records behind those costs — counts of tokens, characters and seconds, holding no content.
  • A content_deleted_at stamp, so the call reads as erased rather than as empty.
The line is what an invoice needs, not what a debugger would like. Everything a bill is built from stays; everything else goes.

​
The deadline is fixed when the call ends

Each call’s purge is scheduled once, at the moment it ends, from the policy as it stood then. That has two consequences worth planning around:
  • Changing the policy never moves a deadline that already exists. Shortening it to 30 days does not retroactively destroy the calls you already have; lengthening it does not rescue calls already scheduled to go.
  • A new policy applies to calls from that point on. If you need old calls gone sooner, delete them yourself.
A workspace with no policy schedules nothing at all, which is why the default costs nothing.

​
Deleting a call yourself

DELETE /v1/calls/{session_id} runs exactly the same purge, immediately.
It answers 202: the erase is scheduled rather than done inline, which is what stops a request covering two hundred thousand calls arriving as two hundred thousand synchronous deletes. It normally completes within seconds. Editors and admins can do it. A call still in progress is refused with 409 — it is still writing, and finalizing would put the content back on a row already marked erased. Ask again once it has ended.
This cannot be undone, and it takes the transcript with the audio. To delete only the audio and keep the transcript, use DELETE /v1/calls/{session_id}/recording instead — that one is immediate and narrow. See recordings.

​
Erasing a person

A conversation is deleted once every call in it has been purged, and the contact record goes with it once they have no conversations left. That is usually what you want when someone asks to be erased, and it has a side effect to know about: the contact’s accumulated userdata is the agent’s cross-call memory. Purge a person’s last conversation and the next call from that number meets an agent that has never heard of them — no name, no history, no preferences. Correct, and worth saying out loud before someone discovers it in a support ticket. See conversation memory. Every voice and video agent records by default — one stereo file per call, the caller on one channel and the agent on the other. recording.enabled on the agent turns it off. There is no per-call switch and no per-number setting: recording is a property of the agent that answers, and on a call with handoffs it follows whichever agent is currently speaking. Whether callers are told is a separate setting, recording.consent, and it is off by default. Set it to disclosure and three things change: the agent’s greeting must literally contain {{consent.notice}} — validated at publish — the agent gains a tool that stops the recording if the caller objects, and a realtime agent is refused the setting outright, because a speech-to-speech model may reword a scripted line and a disclosure has to be said word for word. The default notice is “This call is recorded for quality and training purposes.”, and you can replace it. Recording law differs by jurisdiction and by who is on the call, and Talqing does not decide it for you. The default is on; whether that is lawful where your callers are is yours to determine. Full detail is in call recording.

​
Where your data lives

Everything a workspace owns is regional. A call placed against https://api.in.talqing.com is stored in India; one placed against https://api.us.talqing.com is stored in the United States. Recordings, transcripts and images are held in object storage in the region the call belongs to. Nothing is copied or replicated between regions, and no view in the product merges them. Your identity, your workspace membership and your payments are not regional. See regions.

​
What BYOK means for processing

Talqing holds no provider credentials of its own. Your agent’s language model, speech-to-text, text-to-speech and avatar all run on your accounts, on your keys. The practical privacy consequence: the conversation content those providers see is being processed under your own agreement with them — your terms, your data-processing arrangements, your data-retention and training settings in their console. If you have turned off training on your OpenAI account, that setting is the one in force. Talqing has no separate relationship with those providers on your behalf and no ability to override what you have configured with them. What passes through Talqing is what the platform itself needs: the audio and text of the call while it is running, and whatever is stored afterwards under the policy above. Talqing does use its own model keys for three things, none of which is an agent run: the CoPilots in the editors, the knowledge-base builder, and the voice and avatar galleries before you have stored a key for that provider. See provider keys.

​
What we do not claim

We do not hold a compliance certification, and this page will not imply one. Talqing has no published certification, audit report or compliance programme, and nothing here should be read as a legal assurance about any particular regime. What is documented above is what the software does — the settings, what a purge deletes, what survives it, and where the data sits — so you can evaluate it against your own obligations. If you need something in writing for a procurement or legal review, email hello@talqing.com and ask.

Call recording

Defaults, consent, stereo files, and deleting audio on its own.

Regions

Where each region’s data lives, and what never leaves it.