SGM Context

A context is the assembled, consumable object produced from one or more retrieved memory records. Where a memory record is the stored unit, a context is the delivery unit — what one tool hands to another.

Structure

{
  "id": "ctx_01HZX...",
  "scope": { "project": "myproject" },
  "records": [ /* memory records */ ],
  "assembled_at": "2026-10-10T19:00:00Z",
  "provenance": { "agent": "build-agent", "session": "..." },
  "summary": "Optional human-readable digest of the included records."
}

Field	Required	Meaning
`id`	yes	Context identity
`scope`	yes	The scope this context was assembled within
`records`	yes	The retrieved memory records
`assembled_at`	yes	When the context was built
`provenance`	yes	Who assembled it (see provenance.md)
`summary`	no	Optional digest

Normative structure: context.schema.json.

Semantics

Records are not copied. A context references records; it does not become a second source of truth. If a record changes, a freshly-assembled context reflects the change.
Scope is carried. Every context records the scope it was assembled in, so a consumer never has to guess whether cross-scope data leaked in.
Provenance is carried. The context inherits provenance from its records and adds its own assembly provenance.
Ordering is implementation-defined unless the query specified an explicit ordering.

What a context is for

A context is the object that crosses the boundary between tools:

Agent A ──store──► SGM ──retrieve──► context ──consume──► Agent B

Because every context carries its scope and provenance, Agent B can trust where the memory came from without re-deriving it — this is the core interoperability guarantee SGM provides (see handoffs.md for the cross-tool case).

Empty contexts

An assembly that matches no records returns a context with an empty records array — not an error. records: [] is a valid, meaningful context (“nothing relevant was found”).