Response contract (never a bare list):
    - {"status": "queued", "task_id", "submitted_at", "queue_depth", "estimated_wait_s"}
      — infer=true path; extraction runs in background, poll memory_task_status.
    - {"status": "stored", "memory_ids": [...], "results": [...]}
      — synchronous path; empty memory_ids carries "reason": "no_new_facts".
    - {"error": ...} — failure.
    

    Returns immediately with {"status": "queued", task_id, pages,
    chunks_estimate, estimated_wait_s} — a large document takes many
    minutes; poll memory_task_status(task_id) for chunks_done progress and
    the final memory_ids. Re-submitting the same file returns
    {"status": "already_ingested"} unless force=true. There is NO
    synchronous fallback: if the queue is unavailable the call errors.
    

    States: pending | processing | done | failed_retryable | dead.
    When done, ``result.memory_ids`` lists the memories created/updated
    (fetch them with get_memory) and ``result.events`` includes any
    SUPERSEDED markings. ``last_error`` explains failed/dead tasks.
    

    ASYNCHRONOUS by default (when MEM0_ASYNC_INGEST != false): validates the
    memory exists, then returns {"status": "queued", "task_id", ...} immediately
    while the re-embed + metadata re-classification (a slow llama3.1:8b call) run in
    the background worker — so the call never times out the client. Poll
    memory_task_status(task_id) for the result (memory_id / UPDATE event);
    memory_history(memory_id) shows the old-vs-new diff. An identical re-submit
    while the job is still active returns the same task_id (no double-apply). Set
    MEM0_ASYNC_INGEST=false for the synchronous path.
    

    NEVER calls memory.delete_all() — uses safe bulk-delete instead.
    

    Functionally equivalent to delete_all_memories in self-hosted mode.