How context groups keep answers scoped
One giant knowledge pile pulls unrelated threads into every answer. Context groups scope retrieval so finance questions do not return eng debates, and most teams already think in these buckets.
Ask a finance question in a tool that searches Office, GitHub, and Xero at once. You get an eng thread about service boundaries, a portfolio memo from a deal that closed two years ago, and a conflicts note from counsel that has nothing to do with billing. All somewhat relevant by embedding distance. None actually answering the question.
That is what happens when a company brain has no walls.
Why one pile fails
Unified search treats one retrieval space as a feature. For "where is the Q3 PRD," fine. For "what did we decide about EU payment routing after the Stripe migration," it is a bug.
Different functions ask different questions with different vocabularies. Operator handoffs, portfolio judgment, partner precedents, and employee succession memory are different kinds of knowledge. Mixing them in one index produces noisy answers, low trust, and a return to pinging whoever remembers which thread actually mattered.
Context groups are scoped baskets. When someone Ask inside a group, retrieval stays in that bucket.
Four groups on the Horizon demo
The Horizon Logistics demo runs five brains, each with its own groups. Four show the pattern.
Operator holds day-to-day judgment: who owns a billing dispute, what changed after the 2022 reorg, when carrier renewals need founder review in Q4. Questions here should not pull portfolio IC notes or Morrison conflict precedents.
Portfolios holds investment judgment: pass reasons on Coastal Freight, the 45% concentration threshold, diligence gaps someone flagged in committee. A question about deal structure should not return ops handoff threads from Operator.
Partner holds firm judgment for counsel: Morrison's four-business-hour conflict intake SLA, carve-out templates that cite prior matter IDs, the success fee cap from Meridian. Legal Ask here should not return eng billing debates from Employee.
Employee holds succession and ownership: who holds the Stripe relationship, deploy rights on the payments service, knowledge that walks out when someone leaves. Questions about billing architecture should not pull unrelated partner memos.
Same company. Same Office workspace connected. Different questions, different walls.
What goes in each group
Groups map to how teams already partition knowledge, even when no software enforced it.
Operator: handoff paths, regional ops changes, dispute resolution, process exceptions.
Portfolios: pass reasons, concentration thresholds, diligence gaps, dissent arguments.
Partner: conflicts precedents, fee caps, matter IDs, carve-out templates.
Employee: who owns vendor relationships, deploy rights, knowledge tied to specific people.
Decisions land in one primary group. Ask respects that boundary unless someone explicitly searches across groups.
Private, team, and published layers
Not everything belongs in a shared company basket on day one.
Personal groups stay private until someone publishes upward. A corp dev associate might capture IC judgment in a private Portfolios group, then publish selected decisions to the team layer after the memo is filed.
That opt-in model matters for adoption. If everything shared by default, people stop saving. If private groups stay private until deliberately shared, capture feels safe.
What a company brain captures covers the privacy and employer visibility model.
What breaks without groups
Noise. A question about Stripe vs Braintree for EU billing returns a portfolio payment setup thread because both mention "Stripe."
Wrong confidence. The system surfaces an old eng debate with the same keywords as the current decision, with no signal about which decision superseded which.
Trust collapse. If Ask is wrong often enough, people go back to Slack pings. The brain becomes shelfware.
Groups do not fix bad capture. They fix bad retrieval when capture is good.
How groups interact with the capture loop
Connect Office, GitHub, and Xero surface candidates from connected tools. On confirm, the saver picks a group (or accepts a suggested one). The decision inherits group membership. Future Ask queries inside that group only.
Suggested groups can come from channel context, decision kind, or prior patterns. The human still confirms. Automation suggests; it does not assign silently.
For the full capture loop, see What is a company brain?.
When to split vs merge groups
Split when the audiences differ (legal vs ops), the vocabulary overlaps enough to confuse retrieval (billing service vs billing disputes), or confidentiality boundaries require it (deal team vs integration team pre-announcement).
Merge when a team is small enough that one basket stays precise, or when two groups never produce cross-contamination in practice.
Most teams start with three to five groups and split when Ask noise appears, not before.
Try the Horizon Logistics demo to switch between Operator, Portfolios, Partner, and Employee brains and see how scoped Ask changes the answers.
For the team product layer with five brains, see Company Brain for teams.
Connect in the desktop app
OAuth link and sync scope per tool—signals feed Capture, nothing saves until you confirm.
Join the waitlist
Company Brain is in early access. Leave your email and we'll invite you when your spot opens.
Talk to us about a team rollout