Python memory API

The embedded engine is useful without a hosted model. Run the complete isolated quickstart first.

Main operations

Call	Purpose
`Memory(path)`	Open canonical memory and its normal derived vector sidecar
`remember(subject_id, message, session_id=...)`	Admit a sourced memory through policy
`recall(subject_id, query, limit=...)`	Return matching eligible records
`verify(subject_id)`	Check retained integrity
`close()`	Flush/close owned resources

Procedure review authority (2.3.6)

An embedded application that can approve typed procedures must configure its review principals when it opens Memory. A plain actor string is audit text, not authority:

from atmem import Memory
from atmem.extract import ReviewService

memory = Memory(
    "memories.db",
    review_authorities=(
        {
            "principal_id": "owner:alice",
            "subject_id": "alice",
            "agent_id": "assistant",
            "workspace_id": "personal",
            "scopes": ("procedure",),
            "assurance": "host_authenticated",
        },
    ),
)
authorization = memory.issue_review_authorization(
    "owner:alice", scopes=("procedure",)
)
ReviewService(memory).decide(
    proposal_id,
    "approve",
    actor="display-only",
    authorization=authorization,
)

The issued value is bound to that Memory instance, configured principal, permission, subject, agent, and workspace. Caller-created values, widened scopes, altered fields, and cross-instance reuse fail closed. The signed-in local dashboard performs this mapping for an Administrator account.

Memory.submit_proposal also refuses secret-bearing or explicitly excluded semantic payloads before creating active or quarantined records. Captured source evidence remains subject to its separate evidence access and retention policy.

A Python process using the embedded API owns its storage access. This is not the same authentication boundary as a dashboard account or remote HTTP client. Never turn user-supplied subject IDs into authorization without checking them.

Retrieval versus injection

Calling recall does not prove that context was sent to a model. Use a host adapter for preparation, exact delivery confirmation and lifecycle evidence. Use retrieval quality to understand support checks and the opt-in fusion profile.

HTTP client

atmem.client.AtMemClient accepts an endpoint and local credential. It is a different interface from the embedded Memory class. See HTTP API before exposing any service beyond loopback.

← CLI MCP profiles →