Memoripy v4 API reference

August 17, 2026 ยท View on GitHub

Memory

A small facade for the common workflow.

Memory(path=None, **client_kwargs)

Methods:

  • capture(text=None, **kwargs)
  • recall(query, **kwargs)
  • search(query, **kwargs)
  • write(key=..., value=..., **kwargs)
  • explain(memory_id)
  • correct(memory_id, value, **kwargs)
  • forget(memory_id, **kwargs)
  • audit()

MemoryClient

Construction

MemoryClient(
    repository=None,
    chat_model=None,
    embedding_model=None,
    extractor=None,
    pipeline=None,
)

MemoryClient.from_path(path, **kwargs)

Capture and writing

client.capture(
    messages=None,
    events=None,
    items=None,
    user_id=None,
    agent_id=None,
    run_id=None,
    project_id=None,
    organization_id=None,
    namespace=None,
    idempotency_key=None,
)

Returns evidence IDs, created and updated records, pending records, quarantined candidates, rejected candidates, admission decisions, and projection status.

client.write(
    kind,
    key,
    value,
    summary=None,
    subject=None,
    metadata=None,
    tags=None,
    entity_names=None,
    layer="semantic",
    observed_at=None,
    durability="durable",
    trust_level="authoritative",
    valid_from=None,
    valid_to=None,
    **scope,
)

client.search(
    query,
    limit=5,
    filters=None,
    include_trace=False,
    include_historical=False,
    as_of=None,
    expand_scope=True,
    track_usage=True,
    **scope,
)

Search results contain:

  • memory
  • score
  • rank_breakdown
  • retrieval_receipt
  • evidence
  • projection_status

Context

client.context.build(
    query=None,
    messages=None,
    limit=8,
    max_tokens=480,
    include_debug=False,
    include_trace=False,
    include_historical=False,
    as_of=None,
    context_policy="compact",
    **scope,
)

Returns a ContextPack.

Record operations

  • get(memory_id=...)
  • get_all(filters=None, **scope)
  • update(memory_id=..., data=..., idempotency_key=None)
  • correct(memory_id=..., value=..., reason=None, valid_from=None, idempotency_key=None)
  • delete(memory_id=..., idempotency_key=None)
  • forget(memory_id=..., idempotency_key=None)
  • delete_all(filters=None, idempotency_key=None, **scope)
  • history(memory_id=...)
  • explain(memory_id=...)

Quality and lifecycle

  • audit()
  • feedback(memory_id=..., outcome=...)
  • maintenance.consolidate(...)
  • recover() for file repositories

Supported feedback outcomes:

  • included
  • used
  • success
  • confirmed
  • corrected
  • rejected
  • failure

Import and export

  • export()
  • import_(payload, mode="merge" | "replace", idempotency_key=None)

AsyncMemoryClient

The asynchronous client mirrors MemoryClient and runs synchronous repository operations through asyncio.to_thread.

Configuration

AdmissionConfig

Controls confidence thresholds, assistant semantic writes, generated-summary writes, sensitive-data quarantine, external-instruction quarantine, evidence-span requirements, and admission-log retention.

RetrievalConfig

Controls reciprocal-rank fusion and lane weights for lexical, semantic, exact, entity, temporal, authority, activation, and policy retrieval.

BrainConfig

Controls attention mode, working-memory size, decay, dormancy, consolidation, and utility weighting.

MemoryPipelineConfig

Combines extractor, admission policy, reconciler, reranker, asset processor, brain, retrieval, and trace defaults.

Service

  • MemoryService
  • serve_http(...)
  • create_fastapi_app(...)

The service is a local-development surface unless the caller supplies deployment authentication, authorization, tenancy, and operational controls.