Snapshots & data retention
Snapshots, run history, store rows and AI memory are kept until you delete them; there is no 30-day purge. Covers where display truncates and deletion.
The short version: MadMax keeps your history. There is no
quiet 30-day purge, and no background job that deletes runs, rows
or documents on a timer. The only scheduled sweepers do housekeeping
of a different kind: they mark a run failed when its job vanished
mid-flight, and they clear steps still bound to a connection you
deleted.
Kept indefinitely
- Snapshots: every activation and every Publish edits cuts an immutable snapshot; all of them stay, forever, as the audit trail of what was live when. Snapshot history in the canvas topbar lists them newest-first and can load any one back onto the canvas.
- Run history: runs and their per-step inputs/outputs are not pruned. The runs page, the REST API and MCP all read the same rows.
- Store rows: data stores hold rows until a workflow overwrites them (upserts replace; there's no row history) or you delete the store. Each store has a row cap (100,000 unless the store was created with its own); when it's full, saves fail with a clear error rather than evicting old rows.
- AI memory: conversation turns persist until you clear them; the window limits what an agent reads, not what's stored. Older turns beyond the window are folded into a rolling session summary the agent also sees, so a long-running session keeps its gist as well as its raw turns.
- Knowledge bases: uploaded documents and their chunks stay
until you delete the document (a
knowledge_writestep with the same source name replaces it) or the whole base. - Attachments and uploads: files received by email triggers and upload pages are stored in object storage under the key the payload carries. There is no timed cleanup of those objects either.
Where display truncates
Storage and display differ in a few places, and the totals always come from the full data:
- Long loops show the most recent 50 iterations on the run page, with the OK/failed counts computed over every iteration.
- The runs list pages in batches of 25; scroll for older runs or filter by status.
- Big step outputs are clipped when stored: web request responses are capped at 4 MB and inbound email bodies at 512 KB, so one monster payload can't bloat a run.
Deletion is explicit
Data leaves MadMax when you delete it: a store (typing its slug to confirm), a memory (typing its name to confirm) or one of its sessions (one click clears the transcript), a knowledge base (typing its name), a workflow, a connection, or an archived workspace. Deleting a workflow keeps nothing running: its triggers tear down with it and it leaves the index. Deleting a connection unbinds the steps that used it and turns off the live workflows that depended on it, so nothing keeps firing into a credential that no longer exists.
Account deletion is on the profile tab of Settings. Workspace archiving is owner-only, cancels any subscription immediately, and preserves content on disk in an inaccessible state; it is not self-serve reversible, so treat it as the shutdown it is.