Enterprise Server

Configuration reference

The settings an Enterprise Server reads, grouped by what they govern: the ones an operator is likely to change. The server declares every KNWLGE_* variable in one typed schema, and Settings → Configuration lists all of them with their current state.

How settings are read

The native server (knwlge-enterprise start) builds its environment from config.json, which setup wrote, plus any value saved in the console. A value set in the process's own environment wins over both.
Docker Compose and Helm read the environment of their definitions; the console shows the values but cannot change them.
Unknown, invalid or retired settings are logged as warnings by name and the server starts anyway; KNWLGE_CONFIG_STRICT=true makes them fail the start instead. Values are never logged.

In the tables, Console says whether an admin can change a setting in Settings → Configuration (taking effect at the next restart), or why not.

Knwlge Global

Setting	Default	Console	What it does
`KNWLGE_GLOBAL_URL`	—	No — setup	Knwlge Global's address, https://api.knwlge.com. Turns on enrollment, Global sign-in, the heartbeat and usage export.
`KNWLGE_GLOBAL_SERVER_URL`	—	No — setup	This server's canonical public address. Sent at enrollment; every token the server accepts must be for it.
`KNWLGE_GLOBAL_ENROLLMENT_TOKEN`	—	No — secret	An enrollment key. Enrolls the first start; left set, it re-enrolls the server on its own if Global refuses it while the key is still active.
`KNWLGE_GLOBAL_STATUS_INTERVAL_SECONDS`	`90`	No — setup	Seconds between status reports to Global, 60 to 300, with a little jitter.
`KNWLGE_GLOBAL_STATUS_DRY_RUN`	`false`	No — setup	Log each status report instead of sending it — to see exactly what would leave.
`KNWLGE_GLOBAL_HTTP_TIMEOUT_MS`	`10000`	No — setup	Timeout for every call from this server to Global.

Clients and sign-in

Setting	Default	Console	What it does
`KNWLGE_ALLOWED_CLIENTS`	`codex,claude,copilot,cursor`	Settings → Enterprise Server	The AI clients people may use: codex, claude, copilot, cursor, grok. A list saved in the console wins until it is reset.
`KNWLGE_AUTH_PROVIDERS`	`knwlge`	Settings → Sign-in	Sign-in methods accepted: knwlge, microsoft, github (knwlge is always included). The console's settings, with email domains, win until reset.
`KNWLGE_MULTI_TEAM_READS`	`true`	Yes	Someone in several teams reads team memories across all of them. Writes always go to the active team.
`KNWLGE_CONTEXT_KILL_SWITCH`	—	Yes	true stops serving context to every assistant, at once. The server keeps running.

Memory and review

Setting	Default	Console	What it does
`KNWLGE_MEMORY_ACTIVATION_POLICY`	`immediate`	Yes	How shared memories take effect: immediate (with history), tiered (needs verified provenance, else review), review_all.
`KNWLGE_MEMORY_ADMIN_SELF_APPROVAL`	`false`	Yes	Lets an admin approve a candidate or upload they added themselves. Each one is audited.
`KNWLGE_MEMORY_REVIEW_SEPARATION`	`none`	Yes	How broadly the rule that someone else reviews applies: none, broad or all.
`KNWLGE_MEMORY_CANDIDATE_EXPIRY_DAYS`	`30`	Yes	Days an unreviewed candidate waits before it lapses.
`KNWLGE_MEMORY_EXTRACTION_CONFIDENCE_FLOOR`	`0.5`	Yes	The least confidence, 0 to 1, for a memory extracted from a session to be kept.
`KNWLGE_MEMORY_CONFLICT_SIMILARITY`	`0.92`	Yes	Above this similarity a new memory is treated as conflicting with an existing one.
`KNWLGE_MEMORY_RETENTION_DAYS_PERSONAL`	`365`	Yes	Days personal memories are kept.
`KNWLGE_MEMORY_RETENTION_DAYS_REPO`	`0`	Yes	Days repository memories are kept; 0 keeps them until superseded.

Capture and privacy

The capture policy itself is set in Settings → Session capture, not in the environment.

Setting	Default	Console	What it does
`KNWLGE_PII_MODE`	—	Yes	How personal data in captured content is handled on the server: off, mask or block.
`KNWLGE_LLM_REQUIRE_REDACTION`	`true`	No — deployment	Redaction must run before any LLM call. A data-protection guard; only the deployment can turn it off.
`KNWLGE_RETENTION_PURGE_SCHEDULED_ENABLED`	`false`	Yes	Runs the retention purge on a schedule: old session digests and personal memories past their retention.
`KNWLGE_RETENTION_PURGE_DRY_RUN`	`true`	Yes	Report what the retention purge would remove, without removing it.

LLM assistance

Off by default: a default install makes no LLM calls. With a provider key saved in Settings → Provider keys and a call limit above 0, the server's jobs use the LLM for summaries, catalog descriptions and extraction — each call redacted first.

Setting	Default	Console	What it does
`KNWLGE_LLM_PROVIDER`	—	Yes	anthropic or openai-compatible.
`KNWLGE_LLM_MODEL`	—	Yes	The model id for extraction and summaries.
`KNWLGE_LLM_MAX_CALLS_PER_JOB`	—	Yes	LLM calls a job may make; unset or 0 keeps LLM assistance off.
`KNWLGE_LLM_TIMEOUT_MS`	`15000`	Yes	Timeout for each LLM call.
`KNWLGE_CATALOG_SUMMARY_MAX_PER_RUN`	`25`	Yes	Catalog entries the LLM describes after each index; 0 turns it off.
`KNWLGE_CATALOG_SESSION_SUMMARY_MAX`	`5`	Yes	Catalog entries re-described from what one session said about them; 0 turns it off.
`KNWLGE_ENGINE_INDEX_LLM_BUDGET`	`200`	Yes	LLM calls an index pass may make.

Embeddings

A configuration saved in Settings → Embedding provider is authoritative; these are the environment's fallback. The index was built with them, so most are fixed at setup.

Setting	Default	Console	What it does
`KNWLGE_EMBEDDING_PROVIDER`	—	No — setup	openai or http (any OpenAI-compatible or self-hosted endpoint).
`KNWLGE_EMBEDDING_MODEL`	—	No — setup	The embedding model id.
`KNWLGE_EMBEDDING_DIMENSIONS`	—	No — setup	Vector dimensions, for http providers.
`KNWLGE_EMBEDDING_ENDPOINT`	—	No — setup	The endpoint, for http providers.
`KNWLGE_EMBEDDING_API_KEY`	—	No — secret	The provider's key.
`KNWLGE_EMBEDDING_TIMEOUT_MS`	—	Yes	Timeout for each embedding request.

GitHub and Azure DevOps

Saving the App or the organization in Settings wins over these; container installs may keep provisioning them this way.

Setting	Default	Console	What it does
`KNWLGE_GITHUB_APP_ID`	—	Settings → GitHub App	The GitHub App the server verifies installations and repositories with.
`KNWLGE_GITHUB_PRIVATE_KEY_PATH`	—	No — deployment	Path to the App's private key (PEM), readable by the server.
`KNWLGE_GITHUB_WEBHOOK_SECRET`	—	No — secret	Verifies GitHub's webhook signatures.
`KNWLGE_GITHUB_SYNC_DAEMON_ENABLED`	`false`	Yes	Runs the GitHub sync loop inside the worker.
`KNWLGE_ADO_ORG_URL`	—	Settings → Azure DevOps	https://dev.azure.com/<organization>; turns on Azure DevOps.
`KNWLGE_ADO_AUTH_MODE`	—	Settings → Azure DevOps	pat or service_principal.
`KNWLGE_ADO_SYNC_DAEMON_ENABLED`	`false`	Yes	Runs the Azure DevOps sync loop inside the worker.
`KNWLGE_CONNECTOR_SYNC_MAX_PAGES`	`20`	Yes	Pages a connector sync pulls per pass.

Limits and network

Setting	Default	Console	What it does
`KNWLGE_IDENTITY_RATE_LIMIT_PER_MINUTE`	`240`	Yes	Requests a minute from one person.
`KNWLGE_ORG_RATE_LIMIT_PER_MINUTE`	`20000`	Yes	Requests a minute from the whole organization.
`KNWLGE_PROVIDER_ALLOWED_HOSTS`	—	No — setup	The embedding and LLM hosts the server may call; nothing else is reached.
`KNWLGE_PROVIDER_MAX_CALLS_PER_JOB`	`10000`	Yes	Provider calls a job may make.
`KNWLGE_TRUSTED_PROXY_HOPS`	—	No — deployment	Proxies in front of the server whose forwarded headers are trusted, 1 to 4 (or KNWLGE_TRUSTED_PROXY_CIDRS).
`KNWLGE_API_INSTANCE_COUNT`	`1`	No — deployment	API instances; more than one needs Redis for rate limits.
`KNWLGE_METRICS_TOKEN`	—	No — secret	Bearer token required on /metrics; unset leaves it open.

Audit, alerts and retention

Setting	Default	Console	What it does
`KNWLGE_AUDIT_RETENTION_DAYS`	`365`	Yes	Days audit events are kept, 30 to 3650.
`KNWLGE_AUDIT_RETENTION_SCHEDULED_ENABLED`	`false`	Yes	Prunes audit events older than that on a schedule.
`KNWLGE_AUDIT_SIEM_SHIP_ENABLED`	`false`	Yes	Ships audit events to your SIEM continuously.
`KNWLGE_AUDIT_SIEM_ENDPOINT`	—	Yes	Where they go; the bearer token and signing secret are set in the deployment.
`KNWLGE_ALERT_WEBHOOK_URL`	—	Yes	Where the Alerts page sends alerts, signed with KNWLGE_ALERT_WEBHOOK_SECRET. Unset: alerts can be acknowledged but not sent.
`KNWLGE_OTEL_ENABLED`	`false`	Yes	OpenTelemetry traces and metrics.
`KNWLGE_CONFIG_STRICT`	`false`	No — deployment	Fail the start on unknown, invalid or retired settings, instead of warning.