Skip to main content
Company Brain is the durable knowledge layer for Army of Agents. It gives Commander and agents useful context without turning every note, transcript, or one-off correction into permanent truth.

​
When to use this page

Use Company Brain when you want to:
  • preserve company identity, operating rules, or domain knowledge
  • approve or reject memory suggestions from agents
  • decide what context an agent should see during a task
  • clean up stale or incorrectly scoped knowledge
  • understand why Commander or an agent did or did not use a fact

​
How Company Brain works

Memory is not just search. It has layers, scopes, ownership, lifecycle state, and approval rules.

​
Memory layers

Agents can suggest memory, but high-trust durable memory requires human approval. This keeps the Company Brain useful instead of letting every agent output become permanent context.

​
Before you approve memory

Check three things:
  1. Truth: is the statement accurate today?
  2. Scope: should it apply company-wide, to a department, to a project, or only to a task chain?
  3. Durability: should the system remember this for future work, or is it temporary context?
A broad, approved memory item can influence many future runs. Prefer narrower scope when the fact only applies to one department, project, customer, or task chain.

​
Review pending memory

1

Open Memory

Go to Memory Explorer and open Pending Review.
2

Inspect the source

Read the suggested item and open the source task, discussion, upload, or agent run when available.
3

Edit the wording

Rewrite the memory as a durable fact or operating rule. Remove one-off phrasing and temporary details.
4

Choose layer and scope

Select the narrowest layer and scope that will help future work.
5

Approve or reject

Approve useful memory, reject noise, and archive stale entries.

​
How to verify it worked

After approval:
  • the item leaves Pending Review
  • it appears in the correct folder, layer, or scoped view
  • semantic search can find it after embeddings are available
  • a future Commander or agent turn can use it when the actor has visibility
The OpenAI key in Settings → Memory is for embeddings only. Discussion extraction and Crew memory extraction run through local CLI paths, not hosted extraction keys.

​
Troubleshooting

Check whether embeddings are configured and whether indexing has completed. Memory Explorer shows warnings when embeddings are unavailable.
Check the actor, visibility, department, project, goal, and task scope. Retrieval is intentionally actor-aware and scope-aware.
Reject one-off suggestions. Durable suggestions should usually reflect repeated patterns or important company facts.

Memory Explorer

Use the Memory UI to search, review, upload, and inspect graph context.

Discussions

Turn unstructured threads into tasks and memory candidates.