Brain usage and limits

Brain reports five meters. Usage reports describe one product; plan allowances apply to the whole organization. Only llm-tokens can block work, and only under blocking enforcement. Prices come from your plan.

Meters

Meter	Unit and aggregation	What it measures
`llm-tokens`	Tokens, summed over the period	Total language-model input plus output.
`llm-input-tokens`	Tokens, summed	Text sent to language models.
`llm-output-tokens`	Tokens, summed	Text returned by language models.
`embedding-tokens`	Tokens, summed	Text converted into search vectors.
`memories-stored`	Items, latest reading	Active memories across all product scopes, measured daily at 04:00 UTC.

Do not add all three language-model meters together. llm-tokens = llm-input-tokens + llm-output-tokens. Use the total or its input/output breakdown; combining them double-counts usage. Reports give totals per meter, never a sum across meters.

memories-stored is a level, not accumulated consumption. Deleting memories lowers the next day’s reading. It is excluded from the daily token ledger because summing daily levels would not describe usage.

Operations and consumption

Operation	Consumption
Add or correct a memory	Embedding tokens and language-model tokens for extraction.
Search memory or the knowledge graph	Embedding tokens for the query.
Daemon ingestion	Embedding and language-model tokens for new or changed material.
Connector batch submission	Embedding and language-model tokens, proportional to the files processed.
Cognition digest or sweep	Language-model tokens.
Delete a memory	None.
Create or restore a snapshot	None; stored vectors are restored without recomputation.

Tokens follow the work performed, not the request count. A daemon run that finds nothing new and performs no processing consumes nothing. See daemon load modes for hot and batch processing.

Usage report

There is no Brain usage screen in the web app. Use brain_usage_report over MCP with product-scoped read access, or GET /brain/usage/report on the authenticated REST API.

Input	Type	Meaning
`days`	Integer, 1–90	Number of UTC days ending today. Defaults to 30.

Response	Meaning
`product.rows`	This product’s daily token ledger, broken down by meter, model, processing mode, and purpose. Each row includes tokens, work-event count, and failed billing-delivery count.
`product.totals`	Those figures aggregated over the report window, separately for each meter.
`platform`	Organization plan and each meter’s monthly limit, amount used, amount remaining, and unlimited status. `null` when unavailable.
`platform_error`	Why organization billing data could not be read; `null` when it was read successfully. Product ledger data still returns.
`note`	Explains that product usage and organization billing have different scopes and need not match.

The ledger answers what this product used. The plan aggregate includes all products in the organization. A difference between them is not, by itself, a reporting error.

When an allowance runs out

Allowances are numeric monthly limits or unlimited, defined by the organization’s plan. Embedding tokens and stored memories are measured and billed but do not block operations.

Enforcement mode is set for the installation, not by the customer:

Mode	Behavior
Off	No allowance check.
Advisory, the default	Checks and records overages; work proceeds.
Blocking	Checks before billable work starts; exhausted `llm-tokens` allowance stops work as below.

Work under blocking enforcement	Result when allowance is exhausted
Add or correct a memory	Refused with a payment-required error naming the limit.
Delete a memory	Unaffected; deletion is never gated.
Daemon run	Skipped for the cycle, finishing `done` with nothing ingested.
Batch submission	Nothing submitted; the result records that billing blocked it.
Scheduled cognition digest or sweep	Silently skipped, with no run record.

Skipped daemon and cognition passes are not queued for later. When allowance is available again, the next scheduled pass can process material accumulated in the meantime.

If the allowance cannot be read, Brain proceeds and records that it could not check. An unavailable billing service therefore does not block the work.

Delivery of usage records

Condition	Behavior
Normal processing	Each measurement is sent to billing as it happens.
Billing delivery fails	Delivery is retried for about 80 minutes, then the record is set aside.
Persistent failed-delivery counts	Report them; the usage report shows counts per day and meter.
Work cannot be attributed to a product	Its usage record is dropped rather than billed to another product.

Was this page helpful?