Knowledge and memory

The hub keeps two stores. They answer different questions.

	Knowledge	Memory
Who sees it	The whole team on the project	Recalled by project, branch and agent
What goes in	Fixes, patterns, decisions, risks	Session context: what was done, what is next
Store	cortex_knowledge_store(title, content)	cortex_memory_store(content)
Recall	cortex_knowledge_search(query)	cortex_memory_search(query)

Both are embedded and searched by meaning. /cs runs both searches at the start of every session, and the hooks refuse edits until they have run. They are for errors and decisions, not for finding code: use the code tools for that.

Knowledge

cortex_knowledge_store(
  title: "SQLite busy errors under parallel tests",
  content: "Cause: … Fix: … Verified by: …",
  tags: ["sqlite", "testing"],
  category: "error_fix",
  files: ["apps/dashboard-api/src/db/client.ts"]
)

Parameter	Values
category	risk, lesson, error_fix, pattern, decision, general
hallType	fact, event, discovery, preference, advice, general (default)
validFrom	ISO date the fact became true. Search with asOf to get only what was valid on a date
files	Repo-relative files or dir/ the note is about. See file notes below
tags, projectId, agentId	Filters and attribution

cortex_knowledge_search takes query, plus tags, projectId, hallType, asOf and limit (default 5). Results marked deprecated carry a warning.

On the dashboard, Knowledge lists every document, groups them by project, creates new ones by hand and exports them as JSON.

When something breaks

1. cortex_knowledge_search(query: "<error message>"): someone may have fixed it.
2. cortex_memory_search(query: "<error context>"): you may have seen it before.
3. Fix it.
4. If the fix was not obvious, store it with cortex_knowledge_store.

File notes

A note stored with files: [...] is linked to those paths. So are the risks a Conductor task reports when it completes. The next agent that is about to read or edit one of those files gets the notes as context, without having to know a search would have found them:

• In Claude Code, the recall-notes.sh hook asks the hub once for every linked path of the project (cortex_file_notes), caches the answer for ten minutes, and shows up to three notes per file, once per conversation and agent. It never blocks. CORTEX_NOTES=off turns it off.
• Code tool results that name a linked file have its notes appended.

cortex_file_notes(repo, files) returns them directly, as JSON keyed by path.

Memory

cortex_memory_store(content: "Moved retries into the client; next: wire the dashboard toggle", branch: "feat/retry")
cortex_memory_search(query: "retry work", branch: "feat/retry")

branch needs projectId. A search on a branch falls back to the whole project when the branch has nothing. cortex_memory_delete(id) removes one entry.

cortex_session_end saves the session summary as memory, which is what the next /cs recalls. A session closed automatically (the agent exited without /ce) is kept out of that recall.