Redactions

Memoturn runs a two-stage redaction pipeline on every recorded turn:

1. Deterministic scan — 25+ regex patterns match API keys, tokens, passwords, private keys, and connection strings. Matched spans are replaced with [REDACTED:type] before the turn is stored.
2. LLM review — an async pass re-reads the redacted turn and flags anything the regexes missed (novel secret formats, credentials in natural language). Results are classified as confirmed, escalated, or dismissed.

Escalated turns appear in the dashboard’s redactions tab and via the list_redaction_reviews MCP tool.

REST API

List redactions

curl "https://YOUR_API_HOST/v1/projects/my-project/redactions?status=escalated" \
  -H "Authorization: Bearer mt_..."

param	type	notes
status	"escalated" | "acknowledged" | "dismissed"	Default "escalated"
since	number?	Start of time window (ms epoch)
tool	string?	Filter by originating tool
limit	number	1–200, default 50

Acknowledge a redaction

Mark a flagged turn as reviewed (keep the redacted content, no further action needed):

curl -X POST https://YOUR_API_HOST/v1/projects/my-project/redactions/TURN_ID/acknowledge \
  -H "Authorization: Bearer mt_..."

Forget a redaction

Delete the turn’s derived memories and mark it as handled:

curl -X POST https://YOUR_API_HOST/v1/projects/my-project/redactions/TURN_ID/forget \
  -H "Authorization: Bearer mt_..."

This soft-deletes all memories derived from the turn, evicts it from the hot cache, and broadcasts memory_forgotten events to subscribed clients. The turn row itself is preserved with its redacted content for audit purposes.

MCP tool

The list_redaction_reviews tool exposes the same data to MCP clients:

list_redaction_reviews({ status: "escalated", limit: 10 })

Use it to surface secret leaks an agent should not have written, without subscribing to broadcasts.