Curating what your channel Claude learns — and the privacy trade-off underneath it.
You know memory is channel-scoped and persistent (Lesson 2). This lesson is about its lifecycle: what goes in, how it's governed, and how to stop it rotting — the same discipline you apply to your own memory notes.
By default a ClaudeTag learns from its own channel's activity. With permission it "can even automatically learn from other Slack channels and data sources."[ref §3] Independent coverage put the flip side plainly: it's "learning your company one Slack message at a time."[TechCrunch] Persistent memory is an asset and a liability in the same breath.
Admins can view, edit, and delete stored memory in Organization settings.[ref] That turns memory from a black box into something you curate, exactly like your MEMORY.md:
| Your memory hygiene | ClaudeTag equivalent |
|---|---|
| Delete memories that turn out wrong | Edit/delete a stale channel memory before it misleads |
| Don't store what the repo already records | Don't let it memorise noise it can re-derive |
| Verify a recalled fact still holds | Treat old channel memory as "true when written" |
Two facts bound the risk: channel memory is kept separate from your Claude.ai history, and it stays scoped to the channels an admin defines.[ref] For a live-trading estate that means: keep secrets out of channel chatter (they'd become memory), and never bridge a sensitive channel outward without a reason (Lesson 4's wall).
Anything you'd refuse to paste into a memory note: live credentials, account passwords, API keys, secrets. Your estate already keeps these in Keychain / a DPAPI store precisely so they never land in plaintext notes — the same rule holds for a channel a Claude is memorising. If a secret must be used, it belongs in a scoped Access-bundle credential, not typed into the thread.
A channel memory from months ago names a now-burned account. Best practice?
Memory is "true when written." Admins can edit/delete it, so curate stale entries — same as pruning a wrong MEMORY.md note. Stored memory is not self-refreshing, and spend caps are unrelated.
Why keep live credentials out of a Claude-watched channel?
What's said in the channel can be remembered — so a pasted secret becomes stored memory. Use scoped Access-bundle credentials instead. Secrets don't affect latency or billing.
You can treat a ClaudeTag's memory as a curated store, not magic: seed it deliberately, prune what's stale, and keep secrets out. It's your existing memory hygiene, pointed at a channel.