Python SDK¶

The myelinate package includes clients and framework adapters. Install it with the extras you need:

pip install myelinate             # client + gateway
pip install "myelinate[maf]"      # + Microsoft Agent Framework adapter
pip install "myelinate[otel]"     # + OpenTelemetry tracing of the gateway itself

Clients¶

Both clients implement the same LearnerClient protocol, so adapters work with either.

`GatewayClient`¶

Talks to a remote gateway over HTTP. This is recommended for teams: knowledge and governance are shared across replicas and frameworks.

from myelinate import GatewayClient

client = GatewayClient(base_url="https://myelinate.internal", api_key="myl_...", timeout=10.0, retries=2)

base_url and api_key default to the MYELINATE_URL and MYELINATE_API_KEY environment variables. Requests that fail with a connection error, 429 or 5xx are retried with backoff. record() attaches an idempotency key, so a retried upload never records an episode twice.

`EmbeddedClient`¶

Runs the whole engine in your process, with no server. It suits notebooks, tests, edge devices and single-process apps.

from myelinate import EmbeddedClient

client = await EmbeddedClient(tenant_id="local").start()   # starts a background worker too
...
await client.aclose()

It uses the same configuration, for example MYELINATE_DATABASE_URL and ANTHROPIC_API_KEY.

Methods¶

Method	Purpose
`await recall(RecallRequest)` → `RecallResponse`	Skills, memories and profile for a task, plus ready-to-inject `instructions`
`await record(EpisodeIn)` → `EpisodeOut`	Record a finished run
`await feedback(FeedbackIn)`	Rate a run, optionally with a correction
`await remember(MemoryIn)` → `list[str]`	Add a memory (reconciled against existing ones)
`await ask(AskIn)` → `AskOut`	Answer a question from learned evidence
`await skill_markdown(agent_id, name)` → `str or None`	A skill as SKILL.md
`await aclose()`	Close connections

The models live in myelinate.models: RecallRequest, EpisodeIn, FeedbackIn, MemoryIn, AskIn, Scope, Message, ToolCall, Outcome and Verification. Their fields are listed in the REST API schemas.

from myelinate.models import EpisodeIn, Message, Outcome, RecallRequest, Scope, Verification

r = await client.recall(RecallRequest(scope=Scope(agent_id="support-bot", user_id="alice"), query=task))
system_prompt = BASE_PROMPT + "\n\n" + r.instructions
...
await client.record(EpisodeIn(
    scope=Scope(agent_id="support-bot", user_id="alice"), recall_id=r.recall_id, task=task,
    messages=[Message(role="user", content=task), Message(role="assistant", content=answer)],
    outcome=Outcome.success, verification=Verification(passed=True, details="order refunded"),
))

Tools for your agent¶

make_tools(client, agent_id, user_id=None) returns two plain async functions with docstrings. Most frameworks accept these directly as tools:

skill_view(name) loads a skill listed under "Other learned skills".
remember(content, kind="fact") saves a fact, preference or lesson.

Adapters¶

Adapter	Import	Guide
`SelfLearningProvider`	`myelinate.adapters.maf`	Microsoft Agent Framework
`SelfLearning` (context manager per turn)	`myelinate.adapters.generic`	LangGraph, CrewAI and others

Both fail open by default: if the gateway is unreachable, your agent runs without learned context and a warning is logged.