Supported composition

Build knowledge from completed agent work

Flow

From an installed repository checkout:

python -m examples.quickstarts.agent_knowledge

Mari consumes observable activity after the host runtime has executed it. A caller-supplied model labels the workflow and phases. Mari checks those labels against the event sequence, then validates any proposed memory mutation.

Completed activity to a reviewable knowledge proposal

"""Turn completed tool activity into a validated, reviewable memory plan."""

from mari_kit.knowledge import (
    MemoryDecision,
    MemoryOperation,
    plan_memory_mutations,
)
from mari_kit.trajectories import parse_trajectory_analysis


def run() -> dict[str, object]:
    events = (
        {"name": "search", "args": {"query": "refund policy"}, "ok": True},
        {"name": "read_document", "args": {"document_id": "refunds"}, "ok": True},
        {"name": "approve_answer", "args": {"answer_id": "answer-7"}, "ok": True},
    )
    analysis = parse_trajectory_analysis(
        events,
        {
            "workflow": "Find the current policy, inspect its source, and approve the answer.",
            "activity": "Answer a policy question",
            "category": "support",
            "intent": "resolve_refund_question",
            "rework": 0,
            "phases": [
                {
                    "name": "research",
                    "family": "discover",
                    "start": 0,
                    "end": 1,
                    "substate": "grounding",
                },
                {
                    "name": "review",
                    "family": "approve",
                    "start": 2,
                    "end": 2,
                    "substate": "complete",
                },
            ],
        },
    )
    candidates = {"procedure:refund-answer": analysis.grounded_workflow}
    plan = plan_memory_mutations(
        {},
        candidates,
        {
            "procedure:refund-answer": MemoryDecision(
                operation=MemoryOperation.ADD,
                reason="successful reviewed run",
            )
        },
    )
    return {
        "analysis": analysis,
        "proposal": plan.writes[0],
        "target_id": plan.writes[0].target_id,
    }


if __name__ == "__main__":
    print(run())

The returned mutation plan contains proposed writes. The host can send those proposals through evidence checks, human review, policy evaluation, and its own storage transaction. The host retains runtime and commit authority.

Input	Mari operation	Output
Completed tool events	parse_trajectory_analysis	Validated workflow, intent, phases, and rework
Existing knowledge and candidates	plan_memory_mutations	Add, update, delete, and no-op proposals
Approved plan	Host transaction	Application-owned stored knowledge

Search observed knowledge

Use conversation knowledge when the goal is retrieving lessons and decisions from messages or visible tool observations. Use trajectory mining for process structure and procedural learning for reviewable procedure candidates. These paths retain distinct evidence and promotion gates.