Seclai

Seclai

Sign inStart for free

Data Retention

View as Markdown

Data Retention

Seclai keeps the data your agents and content sources produce for a set period, then deletes it. Data Retention is where you see the windows you control in one place and change them.

Retention does two jobs at once. It is a privacy control — how long your inputs, outputs and evaluated content exist on our systems — and it is a cost control, because stored data is billed daily once it passes its free period.

Where to Find It

Data Retention is in the left navigation, under the account you currently have selected. Its windows apply to that account, so an organization and your personal account each have their own.

It has three tabs:

  • Agent traces — the execution history of every agent run.
  • Governance evaluations — the stored content from policy tests and retroactive reruns.
  • Per-resource — links to the windows that belong to an individual knowledge base, content source or memory bank.

What Each Window Covers

Agent traces — default 7 days, and the first 7 days are free. Covers, for every run:

  • the run's input and output;
  • each step's input and output;
  • tool-call arguments and results;
  • any uploaded input files;
  • the searchable copy that powers trace search.

Governance evaluations — default keep forever, with the first 30 days free. Covers observational evaluations only — policy tests and retroactive reruns: the excerpt, explanation and retrieved knowledge-base passages each one shows, the redacted content they store, and the raw prompt and response of each model call that completed for one. A draft preview stores no evaluation, and a policy test run against a live screening point is not observational, so their raw prompts follow the 30-day rule below. Evaluations that ran as part of an agent run are part of that run's trace and follow the agent traces window instead — they are not governed twice.

Both accept a preset, a custom number of days, or Forever.

Held source content — default 30 days, set on the Governance evaluations tab. Covers items an Enforce governance policy blocked when a content source imported them. A held item that is not released by review is permanently deleted when the window ends; a source whose own retention is shorter uses that instead. It accepts a preset or a custom number of days, but not Forever, and held items are not charged storage.

Raw model prompts and responses outside a run — a fixed 30 days, not configurable. The AI assistants, the model playground, step and policy tests, and background summarization each record the raw prompt sent to the model and its raw response. A call follows the agent traces window when its run's trace records it, and the governance evaluations window when it completed for a stored observational evaluation. Every other one is deleted by a nightly sweep once it is more than 30 days old, whatever those windows are set to — including a draft preview, a governance check whose model call failed, and an evaluation run on demand against a finished run. Nothing in the app displays the ones outside a run, so no page changes when they go.

Deletion Is Permanent

When a record passes its window, a nightly sweep deletes its content. There is no undo, no grace period and no recovery. Support cannot restore it.

For agent traces specifically, it is worth being precise about what survives:

  • The run stays in the Traces tab, marked expired. Its status, timing, step structure and credit usage remain accurate — those are what the run list, your usage reports and your invoices read.
  • Its content is gone: inputs, outputs, tool-call arguments and results can no longer be viewed, searched or exported. Text recorded alongside that content, such as the error message on a step or a tool call, may no longer be available either. An alert raised for the run keeps its title, type and figures, and the failed step where there was one, but loses its descriptive text, which can quote an error message; an alert email already sent is not recalled. The run keeps its metadata, which for an email-triggered run includes the message's subject and addresses.

Two consequences worth planning around:

  1. Lowering a window acts retroactively. Setting agent traces to 30 days does not only affect runs from now on — the next sweep deletes the content of every run already older than 30 days. The age chart shows you exactly how many that is before you save.
  2. Export before, never after. Exporting a trace keeps a copy outside the window, but only while the content still exists. Once it has aged out there is nothing left to export.

Because of that, shortening a window asks you to confirm and tells you roughly how many records are already past the new age. Lengthening one, or switching to Forever, saves straight away — it cannot lose anything.

If you are unsure, raise the window first and lower it once you have exported what you need.

Deleting an agent deletes its trace content too, on the next nightly sweep, regardless of this window — including Forever. Keep-forever means "keep my agents' traces", not "keep traces of agents I deleted", and the storage they occupy would otherwise be billed indefinitely for an agent that no longer exists. The run rows survive with their status, timing and credits; the content does not. Export first if you need it.

Reading the Age Chart

Each editable tab shows the age distribution of what you currently have stored. Newest is on the left, oldest on the right, and the bar width adapts to how far back your history goes — daily bars for a new account, weekly or monthly once there is more history.

Two labelled lines are drawn on it:

  • free ends — everything to the left of this is inside the free period and is not being charged.
  • deleted after N days — the current window. Everything to the right of it is what the next sweep removes. Move the window left and those bars are what you stop paying for.

Hovering a bar gives that age band's record count and what it costs per day.

On the Agent traces tab, the chart counts only records that are still indexed, because that is what storage bills for. If you lowered the window in the past, let the searchable copies expire, then raised it again, those runs keep their content but can never be billable again — they are counted in a note under the chart instead, since lowering the window would still delete their content permanently.

What Retention Costs

The headline figure is a rate, not a bill: it is what everything currently stored draws per day once past its free period. That is the number that moves when you change the window.

  • Agent traces are billed per embedding of the searchable copy, per day. A run with a lot of content costs more than a short one, and a run that was never indexed costs nothing at any age.
  • Governance evaluations are billed 1 credit per day for each evaluation that stores redacted output, after the free 30 days.

Charges appear on the Credits & Usage page as storage.

Who Can Change It

Any member of the account can see the retention windows — knowing how long your data survives should not require elevated access.

Changing one requires an account owner or administrator, because lowering a window schedules irreversible deletion across the whole account. Viewers see the settings but cannot save them.

Per-Resource Windows

Some resources carry their own window, set on the resource rather than centrally:

  • Content sources — how long ingested content is kept before it is deleted and stops being billed. See Content Sources.
  • Memory banks — how long entries are kept; conversation banks default to 90 days and general banks to keep-forever. See Memory Banks.

The Per-resource tab links to each of them. The Agent Traces content source also shows its window, read-only, and links back here — there is exactly one place to change it, so the two can never disagree.

API and MCP

Retention is readable and settable outside the UI:

  • MCP — get_retention_settings reads the windows and the current daily cost; set_retention_settings changes the agent-trace window, and update_governance_settings changes the governance-evaluation window and the held-source-content window (source_content_hold_days). Both writes require an owner or administrator.
  • Agents cannot change retention. set_retention_settings is refused for calls originating from an agent run, alongside the other destructive tools. An agent that could shorten a retention window could delete the very traces you would use to see what it did.

Runs whose content has been deleted carry a trace_purged_at timestamp in the API, MCP and export payloads. Check that field rather than an empty output — a run that genuinely produced nothing looks identical otherwise.

Next Steps