Key takeaways: "The AI learns your company" is table stakes. Trustworthy memory is harder. Bob separates what is true for the company, what belongs to a channel, and how you work. He shows who taught him each fact and when, updates a stale fact instead of keeping two contradictory answers, writes a receipt when memory changes, and lets you list, edit, or delete it from Slack.
The flaw in one big brain
Here's a real pattern from any team you've ever worked on: the ops lead wants everything as a spreadsheet with pivot-ready columns. The founder wants three bullets and a number. Same report, same company, same "right answer." Different people. Both are correct.
An AI teammate with one undifferentiated memory pool learns a mush of both. Someone says "always give me bullet points" and suddenly everyone gets bullet points. Someone corrects it once in a DM and the correction leaks into a channel it doesn't apply to. The more it "learns your company," the more one person's habits become everyone's defaults. That's not learning. That's averaging.
Three scopes, deliberately
Bob keeps company facts, channel context, and personal preferences separate:
Person memory. How you like things done: your formats, your tone, your shortcuts, which spreadsheet layout you actually open. Private to you: your preferences never restyle a teammate's work.
Org memory. What's true for the whole workspace: the fiscal year starts in February, refunds over $500 need a manager, the weekly report goes to #ops on Fridays. Shared by everyone, because it is everyone's.
Channel memory. What a team decided in one place: the launch plan in #product or the escalation rule in #support. Useful to the people doing that work without turning every local decision into company policy.
When you ask Bob for something, he recalls the relevant company and channel facts plus your way of working, then produces the version of correct that's correct for you. When your teammate asks for the same thing an hour later, they get the version that's correct for them. Same facts, different hands.
The part most memory products skip is accountability. Bob's memory list shows who taught him a fact and links back to when it was learned. If a new instruction contradicts an old one, Bob updates the existing fact and shows the old text crossed out in the memory receipt instead of leaving two truths for a model to choose between. If a judgment call belongs to the teammate who taught the fact, Bob says who to ask instead of bluffing.
Why this matters more as AI teammates spread
The industry is racing toward agents that absorb organizational context, and absorbing context is genuinely valuable. We're building the same muscle. But memory design is trust design. Memory that isn't scoped is memory that leaks: across people, across teams, across the boundary between "my working style" and "company policy." Scoping isn't a feature we added to memory. It's what makes memory safe to use.
FAQ
Can I see or delete what Bob remembers?
Yes. Ask Bob to list memories in Slack. Each fact shows its scope and provenance, and you can edit or delete it from the same flow. Bob also posts a memory receipt when it saves or corrects a fact, so learning is visible instead of silent.
Does my personal memory follow me across workspaces?
No. All memory is scoped inside your workspace. Nothing crosses tenants, ever.
Scheduled jobs: whose preferences do those use?
Schedules retain a server-side owner and use only the memory scopes that run is allowed to see. Company and channel context can guide shared work; one teammate's private preferences do not leak into another person's job.