Call this first in every conversation. Returns a briefing: who the user is and how they want you to work; for project its notes, open tasks and last session (a project with no notes yet says so); for task matching skills and integrations. Without arguments: that profile and the known projects. At most 4,000 characters — longer lists are cut and marked "shortened". bootstrap is the same for older setups; later in the conversation use memory_search for a fact and task_list for the task board. Example: cortex_start(project="Birdhouse", task="slice a bracket for the printer").

Start-of-conversation briefing, identical to cortex_start: who the user is, how they want you to work and the project's notes, tasks and last session. Kept so project instructions that say "call bootstrap" keep working; new setups call cortex_start.

Save a durable, TYPED memory and write it back to the brain. type (REQUIRED) = one of: user (who the user is / preferences) · feedback (how you should work — corrections & confirmed approaches) · project (ongoing work, goals, status — give it a review date) · reference (pointers to resources/URLs/IDs). An untyped memory is REFUSED so the brain stays sorted. Only store DURABLE facts (not session chatter), nothing already in code/git, and NEVER secrets (those go to the vault: secret_set via integration_call). Writing the same title again REPLACES that memory (read it first and fold the old content in); the dedup hint flags related titles to merge.

Read a memory's full content and version by its name — find the name with memory_search. An unknown name returns "No memory named …".

Find memories by words (any order); best matches first — a hit in the title counts most, then tags/project, then the text. Returns name, title and a snippet; memory_read gives the full text.

Find skills for a task: up to 10 lines "name — description", best first (a word in the name counts most, then description, tags, text); no match says so. Call this before specialized work, then skill_load the best match.

Load a skill's full instructions (its whole SKILL.md) by name — find the name with skill_search. An unknown name, or a skill outside your allowed set, gets a message saying so.

List the tasks on the shared agent board — work handed between agents — oldest first, every match (id, status, owner, title). Check it before adding or claiming one (task_add / task_claim via integration_call).

Append a checkpoint to a work session so another LLM/device can resume it — for handing over work in progress; durable facts belong in memory_write. Returns the session id. Old sessions auto-expire after SESSION_RETENTION_DAYS.

Find a tool beyond the core set — devices, services, files, mail, calendars, printing, scanning, SSH, webhooks, scheduling, playbooks, agents and administration. Returns matches best first, each with its parameters (* = required); no match says so. Run one with integration_call. For how-to instructions use skill_search, for stored facts memory_search. Example: integration_search("print a PDF on the office printer").

Run a tool found with integration_search and return its result unchanged — its own errors too; a name you may not call is refused with a pointer to integration_search. Same permissions as calling the tool directly: one that changes or deletes data does so here as well, so confirm such steps as you would there. Example: integration_call(name="print_document", arguments={"printer": "office", "file": "scan.pdf"}).