Configuration reference

Every setting is an environment variable (or a line in .env). Names are MYELINATE_ plus the upper-case field name. Lists take JSON, for example MYELINATE_CORS_ORIGINS='["https://app.example.com"]'.

General

Variable	Type	Default	Description
MYELINATE_ENV	Literal['dev', 'staging', 'prod']	prod	Deployment mode. Secure by default: outside dev the gateway refuses to start without a bootstrap admin key or OIDC. dev with no credentials serves everything as tenant "dev" with all scopes (local use only).
MYELINATE_DATABASE_URL	str	sqlite+aiosqlite:///./myelinate.db	SQLAlchemy async URL. SQLite for dev / single node; PostgreSQL (postgresql+asyncpg://...) for production.

LLM used for reflection, curation, judging and evolution

Variable	Type	Default	Description
MYELINATE_LLM_PROVIDER	Literal['anthropic', 'none']	anthropic	anthropic uses Claude for learning; none disables learning-side LLM calls (recall/record still work).
MYELINATE_ANTHROPIC_API_KEY or ANTHROPIC_API_KEY	SecretStr or None	—	Anthropic API key. Also read from plain ANTHROPIC_API_KEY (env or .env); unset means the SDK's default lookup.
MYELINATE_LLM_MODEL	str	claude-opus-5-5	Model used for learning (not your agent's model).
MYELINATE_LLM_EFFORT	Literal['low', 'medium', 'high', 'xhigh', 'max']	medium	Reasoning effort for learning calls. Lower is cheaper; raise for harder domains.
MYELINATE_LLM_MAX_TOKENS	int	16000	Output token cap per learning call.
MYELINATE_DAILY_TOKEN_BUDGET	int	5000000	Learning-side tokens per tenant per UTC day; jobs are deferred (not failed) when exceeded. 0 = unlimited.

Optional embeddings (any /v1/embeddings-compatible endpoint)

Variable	Type	Default	Description
MYELINATE_EMBEDDINGS_URL	str or None	—	Embeddings endpoint (Voyage, Azure AI, TEI, vLLM, Ollama, LiteLLM...). Unset = lexical retrieval only.
MYELINATE_EMBEDDINGS_MODEL	str	voyage-3.5	Embedding model name sent to the endpoint.
MYELINATE_EMBEDDINGS_API_KEY	SecretStr or None	—	Bearer token for the embeddings endpoint.

Auth

Variable	Type	Default	Description
MYELINATE_BOOTSTRAP_ADMIN_KEY	SecretStr or None	—	Platform-operator key (send with X-Tenant-Id). Use it to mint scoped tenant keys, then vault it.
MYELINATE_OIDC_ISSUER	str or None	—	OIDC issuer URL, checked against the token's iss.
MYELINATE_OIDC_AUDIENCE	str or None	—	Expected token audience (aud).
MYELINATE_OIDC_JWKS_URL	str or None	—	JWKS URL for token signature keys. Setting it enables OIDC authentication.
MYELINATE_OIDC_TENANT_CLAIM	str	tid	Token claim that holds the tenant id (Entra ID: tid).

Recall

Variable	Type	Default	Description
MYELINATE_RECALL_MAX_SKILLS	int	3	Skills injected in full per recall.
MYELINATE_RECALL_MAX_MEMORIES	int	8	Memories injected per recall.
MYELINATE_RECALL_MIN_SCORE	float	0.12	Minimum relevance (0-1) for a skill or memory to be recalled.
MYELINATE_SKILL_INDEX_SIZE	int	10	Further matching skills listed by name only (loaded on demand with skill_view).
MYELINATE_MAX_BULLETS_PER_SKILL_INJECTED	int	40	Cap on rules injected per skill (worst-performing rules are dropped first).
MYELINATE_ASK_FOR_CITATIONS	bool	False	Ask the agent to cite rule ids like [b-1a2b3c4d]: sharper credit assignment, but visible in agent output.
MYELINATE_HOLDOUT_RATE	float	0.05	Share of runs served no learned skills (control group) to measure causal lift. 0 disables.

Learning loop

Variable	Type	Default	Description
MYELINATE_REFLECT_EVERY_N_EPISODES	int	5	Reflect after this many episodes per agent (failures and tool-heavy runs trigger it sooner).
MYELINATE_REFLECT_ON_FAILURE	bool	True	Reflect right after a failed episode.
MYELINATE_REFLECT_MIN_TOOL_CALLS	int	5	Reflect after any episode with at least this many tool calls.
MYELINATE_REFLECT_DEBOUNCE_SECONDS	float	20.0	Wait for an agent to go quiet before reflecting (debounce), in seconds.
MYELINATE_SELF_JUDGE_UNKNOWN_OUTCOMES	bool	True	Use an LLM judge to estimate outcomes when the caller sends none (ReasoningBank-style).
MYELINATE_MIN_JUDGE_CONFIDENCE	float	0.7	Below this judge confidence an episode can teach memories but never change skills.
MYELINATE_MAX_OPS_PER_CHANGESET	int	6	Maximum edit operations per changeset (bounded "textual learning rate").
MYELINATE_DEDUP_SIMILARITY	float	0.85	Cosine similarity at which rules are flagged as possible duplicates (when embeddings are configured).
MYELINATE_DEDUP_JACCARD	float	0.3	Lexical similarity that flags possible duplicates without embeddings; an LLM decides merge or keep.
MYELINATE_CURATE_EVERY_N_EPISODES	int	25	Run library upkeep (dedup, lifecycle, nomination, profiles) every N episodes per agent.

Governance

Variable	Type	Default	Description
MYELINATE_AUTO_APPROVE_MEMORIES	bool	True	Apply new memories without review (still safety-scanned; flagged ones wait for approval).
MYELINATE_AUTO_APPROVE_LOW_RISK	bool	True	Auto-apply purely additive skill changes (new rules / new skills).
MYELINATE_AUTO_APPROVE_MEDIUM_RISK	bool	False	Auto-apply edits to existing rules and duplicate merges. Keep false for agents that act on production.
MYELINATE_AUTO_APPROVE_MIN_EVAL_SCORE	float	0.7	Changesets with an offline eval score below this are never auto-applied.
MYELINATE_MAX_SKILL_CHARS	int	16000	Maximum characters of new text in one changeset (larger ones are flagged).
MYELINATE_MAX_BULLET_WORDS	int	80	Rules longer than this many words get a lint warning.

Canary experiments

Variable	Type	Default	Description
MYELINATE_CANARY_ENABLED	bool	True	Serve changes to proven skills beside the current version on live traffic before applying them.
MYELINATE_CANARY_MIN_SAMPLES	int	20	Uses a skill needs before its changes are canaried, and samples per arm before a verdict.
MYELINATE_CANARY_MAX_SAMPLES	int	300	Samples after which an inconclusive canary keeps the current version.
MYELINATE_CANARY_THRESHOLD	float	0.95	Posterior probability needed to promote (or, mirrored, to reject) a candidate.
MYELINATE_CANARY_FLOOR	float	0.1	Minimum share of traffic each arm always gets (exploration floor).
MYELINATE_CANARY_MAX_DAYS	int	14	Canaries without a verdict after this many days expire and keep the current version.

Evolution (GEPA-style)

Variable	Type	Default	Description
MYELINATE_EVOLVE_MIN_USES	int	10	New uses of a skill required before it can be evolved again.
MYELINATE_EVOLVE_MAX_SUCCESS_RATE	float	0.7	Skills whose posterior success rate is below this are candidates for evolution.
MYELINATE_EVOLVE_ITERATIONS	int	4	Mutation iterations per evolution run.
MYELINATE_EVOLVE_MINIBATCH	int	3	Episodes in each cheap pre-gate minibatch.
MYELINATE_EVOLVE_EVAL_SAMPLE	int	10	Held-out episodes used to score candidates.
MYELINATE_EVOLVE_MIN_IMPROVEMENT	float	0.05	Held-out score gain a candidate needs over the current skill to be proposed.

Lifecycle

Variable	Type	Default	Description
MYELINATE_STALE_AFTER_DAYS	int	14	Days unused before a skill is marked stale (still recallable, ranked lower).
MYELINATE_ARCHIVE_AFTER_DAYS	int	30	Days unused before a skill is archived (not recalled; never deleted). Pinned skills are exempt.
MYELINATE_PROFILE_MAX_CHARS	int	1400	Maximum length of a per-user profile card.
MYELINATE_SHARE_LEARNED_FACTS	bool	False	Share facts learned from one user's runs with all users of the agent (off: they stay user-scoped).

Ops

Variable	Type	Default	Description
MYELINATE_WORKER_ENABLED	bool	True	Run the background job worker inside the gateway process. Set false when running myelinate worker pods.
MYELINATE_WORKER_METRICS_PORT	int	9464	Port for Prometheus metrics served by standalone myelinate worker processes. 0 disables.
MYELINATE_DB_POOL_SIZE	int	10	Database connections per process (PostgreSQL), and extra connections allowed under burst.
MYELINATE_DB_MAX_OVERFLOW	int	20
MYELINATE_WORKER_POLL_SECONDS	float	2.0	Seconds between job-queue polls when idle.
MYELINATE_WORKER_CONCURRENCY	int	4	Jobs a worker runs concurrently.
MYELINATE_JOB_MAX_ATTEMPTS	int	3	Attempts before a job is marked failed.
MYELINATE_RATE_LIMIT_PER_MINUTE	int	600	Requests per minute per authenticated principal (per process).
MYELINATE_IP_RATE_LIMIT_PER_MINUTE	int	3000	Requests per minute per client IP, applied before authentication (per process).
MYELINATE_REDACT_PII	bool	True	Redact emails, phone numbers, card numbers, secrets etc. before anything is stored.
MYELINATE_EPISODE_RETENTION_DAYS	int	90	Episodes, recall logs and finished jobs older than this are removed by myelinate purge.
MYELINATE_AUDIT_RETENTION_DAYS	int	365	Audit-log entries older than this are removed by myelinate purge.
MYELINATE_MAX_REQUEST_BYTES	int	4000000	Maximum request body size in bytes (also enforced for chunked uploads).
MYELINATE_CORS_ORIGINS	list[str]	[]	Browser origins allowed by CORS (JSON list). Empty disables CORS.
MYELINATE_MCP_ALLOWED_HOSTS	list[str]	["localhost:*", "127.0.0.1:*"]	Host headers accepted by the /mcp endpoint (DNS-rebinding protection). Add your public hostname.