SGM Provenance

Memory must not be anonymous text. Provenance is the who/what/when/where metadata that lets a consumer judge whether a record is trustworthy, current, and relevant. It is mandatory on every record — this is a core requirement, not an optional capability.

Why provenance is core

AI-generated memory is uniquely easy to misuse. An agent retrieving a record must be able to tell:

who or what created it — a human decision vs. an automated guess,
when — is this current or months stale,
where it originated — which tool, which session,
whether it was modified — has it been edited since creation,
what it relates to — its scope and links.

Without provenance, an agent cannot distinguish a fact from a stale status from a superseded decision, and will eventually act on the wrong one. So SGM makes provenance a required part of the memory record.

Required provenance fields

Every memory record’s provenance object MUST carry:

Field	Type	Meaning
`agent`	string	Identity of the creator (human, tool, or automated agent)
`created_at`	string (RFC 3339)	When the record was created
`origin`	string	Where it came from (e.g. `cli`, `mcp`, `http`, `import`)
`scope`	object	The scope the record belongs to (see scopes.md)

Optional provenance fields

Field	Type	Meaning
`session`	string	The tool session that created it
`modified`	boolean	Whether the record was edited after creation
`modified_at`	string (RFC 3339)	Last modification time
`modified_by`	string	Who last modified it
`links`	string[]	Related record IDs (see memory.md)
`source`	string	Upstream source, if imported

Semantics

created_at is immutable once set.
If modified is true, modified_at and modified_by SHOULD be present.
scope in provenance MUST match the record’s top-level scope.
Provenance MUST round-trip losslessly through store → retrieve. A consumer reading a record gets the exact provenance it was written with.

Interoperability requirement

Two independent implementations MUST agree on:

the required provenance fields and their meaning,
that provenance is preserved unmodified across retrieval, and
that an unknown provenance field is preserved, not dropped.

They need NOT agree on how provenance is stored or indexed.

Example

"provenance": {
  "agent": "build-agent",
  "session": "20261010-104120",
  "origin": "cli",
  "scope": { "project": "myproject" },
  "created_at": "2026-10-10T18:21:47Z",
  "modified": true,
  "modified_at": "2026-10-10T19:30:00Z",
  "modified_by": "build-agent",
  "source": "session-handoff"
}