What is kept, what leaves, what deleting removes
Three questions a risk review asks first: how long is this kept, who else sees it, and what happens when we delete it. The answers below are the code's, not a policy's - each is what the platform does at the time of writing.
Retention
| Data | Kept |
|---|---|
| Conversation transcripts, attachments | Until deleted. Transcripts are your record of what an agent did on your behalf; the platform never prunes them. |
| Tool-call records (each action's input and output) | 90 days by default, then pruned daily. The floor is 30 days, refused rather than silently widened. |
| Cost, per day | Kept: each closed day is rolled up before its raw records are pruned, and the prune waits for the rollup. |
| Per-turn census behind Health (tool names, counts, tokens, durations) | Kept - facts, never bodies. |
| Audit trail (approvals, credential changes, support access, membership changes) | Permanent. The database refuses to update or delete a row. |
| Files uploaded but never sent | 24 hours, together with the empty draft conversation created to hold them. |
| Caller-supplied credentials | At most one hour, deleted when the run they were supplied to ends. |
What the prune changes on screen is described in Watch and inspect tool calls.
Deleting a conversation
A visible conversation can be deleted by its creator/recorded owner or a domain Owner through the API (DELETE /sessions/:id). Domain membership is still required for domain sessions, including for their creator. Organization Owners/Admins have no general deletion bypass, and personal chats remain private to their creator. Historical sessions from the retired Apps feature retain organization Owner/Admin control. See Roles and permissions. It removes the transcript, every recorded action, and every attachment together with its stored bytes; a flow run started from it keeps its own record, detached.
Two things deliberately survive, and an erasure obligation has to reach them separately:
- A golden eval case saved from that conversation keeps its own copy of the transcript - that copy is what makes the case replayable months later. Delete the case from the agent's Evals page as well.
- Memory the agent saved during the conversation stays in the agent's memory; it is attributed to the conversation but not owned by it. Delete the case from the agent's Memory tab.
The audit trail is untouched by any deletion.
Deleting other things
- A knowledge file on an agent removes its chunks and its bytes; retrieval re-indexes on its own.
- An organization article that was ever published cannot be deleted - its versions are history - but it can be archived. Drafts can be deleted.
- A memory case removes the whole revision chain for that key. A learned rule is removed outright.
- A rating you gave a reply is yours alone to clear; a session's outcome can be withdrawn.
- A person is deactivated, never deleted: their conversations, messages and audit rows remain. See Teams.
What leaves your workspace
On a model call, the provider receives what the turn's context meter shows: the base prompt, the agent's Principles and Procedure, the knowledge passages injected or retrieved for that turn, the rolling summary, the trimmed history, image attachments as images, the definitions of the tools presented, and every tool result. It runs on your own provider key, under your agreement with that provider; there is no fallback to a Yekar.AI key. See Your keys, your model bill.
Knowledge and memory search processes document and article excerpts, saved memories, tool descriptions, and search queries to find relevant matches. Usage records contain counts and character totals rather than the text itself.
Web search uses your query to find titles, links, and snippets from public pages. Browser reads the public URL you provide and returns its page text. Integration tools send the information needed to perform the requested action to the connected service. See Platform security.
Where the bytes live
Records live in PostgreSQL. Files - attachments, uploaded documents, generated media - live in object storage under an opaque key; downloads are short-lived signed links. Credentials are encrypted with a per-organization key that is itself wrapped by a master key, decrypted just in time on the worker, and never readable back through the API: every secret shows as its last four characters and nothing more.
Export
An organization's knowledge - bases, articles with their draft and published bodies, visibility and modules - exports as one JSON snapshot from a domain's knowledge page, and imports the same way. Transcripts and the audit trail are read through the API.
Support access
When Yekar.AI's staff act inside your workspace, it is under a recorded grant with a mode, a stated reason and clocks, visible to you under Settings → Support sessions. A summary email is sent when it ends unless a platform admin with full write impersonation access turned off Inform user for that session. See Platform security.