FAQ

Does Myelinate replace my agent framework? No. Your agent keeps running in its own framework, and Myelinate adds a learning layer around it: recall before a run, record after. It never calls your agent's model.

Which LLM does it use, and does my agent have to use Claude? Learning (reflection, curation, judging, evolution) uses Claude through the Anthropic API (MYELINATE_LLM_MODEL, default claude-opus-5-5). Your agent can use any model. With MYELINATE_LLM_PROVIDER=none, Myelinate still stores and serves knowledge but doesn't learn on its own.

Does it add latency to my agent? recall and record are a few database queries and never call an LLM. All learning happens in background workers.

What does learning cost? It depends on traffic and settings. As a reference point, a full live pipeline run (reflection, curation, memory, evolution with judging, dedup, profile and ask; 31 calls) cost at most about $0.44 on claude-opus-5-5. Cap spend per tenant with MYELINATE_DAILY_TOKEN_BUDGET. See Cost control.

What if it learns something wrong? Several layers guard against it: - Skills only change on reliable evidence. - Risky changes wait for review. - Changes to proven skills are canaried on live traffic. - Rules that keep being marked harmful get evolved. - Every version can be restored with POST /v1/skills/{id}/rollback.

What if Myelinate is down? The adapters fail open: your agent runs without learned context, and a warning is logged.

Can I bring my existing skills? Yes. POST /v1/skills/import accepts SKILL.md files in the Anthropic, Hermes or agentskills.io format, and GET /v1/skills/{name}/SKILL.md exports them back.

Can several agents share knowledge? Agents with the same agent_id share everything. Skills published under agent_id="*" are visible to every agent in the tenant. Only reviewers can publish there (POST /v1/skills or /v1/skills/import with review scope), so one agent can't push knowledge into every other agent.

Is my data sent anywhere? Episodes, after PII redaction, are sent to the configured LLM provider for learning. Nothing else leaves your deployment. See Data and privacy.

How do I know it's actually helping? Look at learning_lift in GET /v1/agents/{agent_id}/insights. It compares success rates between runs that got learned skills and a holdout group that didn't, on comparable tasks, with a credible interval.

Is there a .NET SDK? Not yet. .NET agents (including Microsoft Agent Framework for .NET) can use the MCP endpoint or the REST API.