Thalamus

---
name: thalamus
description: How to use the shared brains of the organisation through the Thalamus gateway (th_brains, th_recall, th_entity, th_get, th_add, th_pending, th_review, th_themes, th_retheme), and what the recall hook injects.
---

# Thalamus: shared brains

The gateway gives you two tools, under your own key. The engine applies your rights on each brain; the gateway records who asked what.

- `th_brains`: the brains you may read, your access (read, write or curate), and whether a key exists for each.
- `th_recall {query, top_k, brains, alternate_queries, deadline_ms}`: one question to every brain you may read, fused into one list. Each result carries a `provenance` line: `[brain] title, id, source, by author`. `skipped` names the brains that did not answer and why. Without `deadline_ms`, the gateway waits for each brain. With `deadline_ms` (at most 60000), it answers at the deadline with the brains that answered, and skips the others with the reason `no answer within N ms`.
- `th_recall` serves first the syntheses of a brain's pyramid, when the brain has one: a result with `level: synthesis` and the provenance `[brain] synthesis: title, id`. A synthesis is written by a model from the documents it cites. Open those documents with `th_descend {brain, knowledge_id}` before you rely on a figure, a date or a name of the synthesis. `syntheses` (default 4, 0 for none) sets how many synthesis chunks come first.
- `th_entity {entity, brains, skip_brains, max_depth, limit}`: one entity across the brains you may read. `known_in` lists the brains that know it, and each neighbor names the brains that serve it. A brain serves only the part of its graph that your key may read. A brain without entities answers nothing: the entities come from the producer of the brain, not from `th_add`.
- `th_get {brain, knowledge_id}`: one document whole. The hook and `th_recall` give extracts, which are cuts: when the answer is not in the extract, read the document with `th_get`, with the brain and the id of its provenance line.
- `th_add {brain, content, title, source_url, document_id, metadata}`: proposes a document into a brain you may write. It waits for a curator; your own document is served at once only when you curate that brain. Write into the brain that the project's `CLAUDE.md` (or `CLAUDE.local.md`) names, unless the user names another one. If none is named, ask the user which brain to use. A project names its brain with a line such as: "Thalamus write brain: `team/acme` (`th_add` with `brain: \"team/acme\"`)."
- `th_pending {brain}` and `th_review {brain, knowledge_id, decision}`: for curators, what waits and the decision (`accepted` or `rejected`).
- In a brain with a convention, `th_add` needs metadata `theme` and `nature`, and a new `document_id` of the form `<nature>:<theme>:<subject>`; the refusal says what is missing. A theme the brain does not know yet files the document under the unsorted theme, with `theme_proposed`. A near copy of a document the brain has is refused, naming it: rewrite that document under its id, or say `replaces` or `distinct_from`. `th_themes {brain}` tells how the brain is filed and which proposed themes are candidates, and names the documents to file: `without_theme_ids`, `unsorted.without_proposal_ids` and the `ids` of each `undeclared` theme (the first 50 of each list in byte order, 200 per call at most; the counts stay whole, and `unnamed_ids` counts the ids the engine sent back as numbers); a curator declares one with `th_add` of `theme:<subject>` and files its documents with `th_retheme {brain, proposed, theme}`.

The hook on each prompt already injects the best chunks of the documents (4 synthesis chunks and 10 passages at most), each with its provenance. A chunk is a passage of the document, not the whole document. When the extract stops before the answer, call `th_get` once; do not call `th_recall` again with the same query, it returns the same extracts. When you need more documents, or a narrower search, call `th_recall` with a precise query, and name `brains` to restrict it. The documents can be in another language than the question: give the question in that language, or in the other languages of the organization, in `alternate_queries` (at most 4). Each variant is searched apart, then fused with the query. On a French question over English documents, an English variant raised the facts found from 81 % to 85 %.

Rules:

1. Quote the provenance line when you use an extract. The reader must know which brain and which document it comes from.
   When the answer carries an `attribution` (a notice that a brain's publisher asks to show, such as "Information provided by Example."), end your answer with that notice, once per text, in the language it comes in. The hook's context ends with it too. Give `th_recall` the user's language in `lang` (en, fr, de, es, it, pt, nl) when you search in another language than the user's.
2. Write through `th_add` only, and tell the user that the document waits for review. Never set `review_status` yourself: the gateway does.
3. A brain in `skipped` with "no key issued" means you have the right but no key yet: ask an administrator for `th-gateway key issue`.