Brain.

Limits

API limits.

The API saves messages and returns current conversation context for a thread.

On this page

Access#

API access is by request. We provide your API URL, Brain ID, application connection and server credential. Request access.

The Python client installs from the Brain repository. There is no separate published package or shared sandbox.

Context selection#

Recall selects current messages by thread ID and returns their original text, versions and citation ranges. It does not generate summaries, infer business schemas, resolve identities across accounts or search across sources.

Semantic memory and generated answers are not enabled. The response reports publication_status: held, semantic_ready: false and accepted_memory: false.

Request bounds#

Input or selectionBound
Original message text1–4,096 UTF-8 bytes; complete text, no NUL.
Message, thread and author IDs1–128 ASCII letters, digits or ._:-; author may be unknown.
Export1–60 messages, at most 1 MiB; stable export ID up to 64 characters.
Recall message limit1–60; default 20.
Recall rendered text budget512–32,000 UTF-8 bytes; default 24,000. Includes evidence headers and separators.
Explicit context selection1–60 distinct current version IDs from one thread; canonical preparation bounds still apply.

A prepared response may still omit messages. Check has_more and omission_reasons. No byte budget is a token guarantee. Latest-saved order is not event chronology; conversation completeness stays unknown.

Access and retention#

Your server must check each user’s thread access. The API credential covers its application connection. Reads check access and return only eligible current versions.

Corrections advance the current version without rewriting history. Withdrawal removes current source eligibility and preserves immutable custody. Withdrawal is not erasure. Contact Brain about retention and erasure requirements before sending sensitive data.

The Python client does not retry automatically. A timeout after dispatch can leave a committed write with an unknown response. Keep exact IDs and content for explicit replay. Error handling reference.

Start with fictional data · Request access