Cloud Brain — Memory-Perfect + Claude Code Always-On

Last updated: 2026-06-09 · Worker 🐝 · answering Sam's "work out the memory + cloud Claude Code" ask.

The goal in one line: talk to Worker 🐝 from your phone/iPad with your laptop closed — and have it be as smart as your laptop session. Two halves, one outcome.

HALF A — MEMORY PERFECT (the brain)

Where we are (good news): #042 already SHIPPED the foundation — the bot AND Claude Code share ONE memory (a cloud D1 catalog, ~257 facts), with retrieval, a nightly "Dreaming" pass that consolidates instead of ballooning, two-way remember→cloud, and a FORGET op. So the memory is already unified + self-cleaning + in the cloud.

What "perfect" still needs (the adaptive layer = the real moat):
1. Correction write-back — when you correct it ("no, it's X"), it records the correction and stops being wrong. Today corrections are ignored (this is GitHub issue #17 / the gap you feel).
2. Nightly distill — compress each day's captures + corrections into clean Profile facts. Self-improving, stays sharp.
3. Cross-device truth — the bot's memory already lives in the cloud (D1). The laggard is Claude Code's LOCAL memory (.claude/, on this one laptop). To be "perfect everywhere," that local memory + skills must also sync to the cloud/repo. (Your queue card #042 + GitHub issue #24 = OneDrive→Drive/cloud.)

Owner: the #042/#043 session, queued right after the current bug fixes (it's the "Adaptive Layer," separate from but built on #042).

HALF B — CLAUDE CODE ALWAYS-ON IN THE CLOUD (the body)

How would you talk to it? (your question)

Through your Telegram bot (or a dedicated channel) — yes. From your end it feels like texting the bot today, EXCEPT the bot would be backed by a full Claude Code session (can run code, deploy, edit files, spawn agents), not just the lighter bot brain. So: same as "remote-controlling your laptop from your phone" right now — minus the need for the laptop to be on. You're on the iPad mini now reaching your laptop; the cloud version reaches a server instead, 24/7.

Where would it live?

A small always-on cloud server — a $5–20/mo Linux VPS (Hetzner / DigitalOcean / Fly.io), running Claude Code with your git repo + the cloud memory (Half A) + the MCP connections. (Or Anthropic's own hosted Claude Code if/when it fits.)

What would it cost — is it worth it?

• Server: ~$5–20/mo. Trivial.
• Claude usage: SAME as now — it runs on your existing Claude plan/API. No double charge. (An always-on agent could use more only if it does heavy background work; it mostly runs when you message it or on schedules.)
• Worth it? For a voice-first, on-the-go operator with multiple businesses — yes. It's literally the "travel the world, laptop closed, operator in my pocket" vision. The money is nothing; the real costs are setup effort + security discipline (below).

How it'd be used / who it's powerful for

• You message it like the bot; it does the Worker 🐝 job — coordinate sessions, manage cards, pull your MIS/market report, run scheduled briefs reliably (always-on = the morning-brief delivery problem disappears).
• Powerful for: you (mobile, voice-first, many plates). Later, scoped, it could run things for Mildred / Chanie / family.

Limitations — what's good, what's bad (honest)

• ✅ Good: always-on, laptop-free, reliable scheduled jobs, reachable from any device.
• ❌ The real limits:
  1. It can't see files that aren't in the cloud. Your big OneDrive archives (Passover ~35GB, the Taxes folder, local screenshots) live on your PC. A cloud Worker can't do deep-file digs on those unless they're synced up (issue #24). So deep local-file work still wants the laptop. Cloud Worker = talk / plan / steer / coordinate / scheduled reports.
  2. Some MCP connectors (Gmail/Drive/Calendar) are interactively logged in and can drop in a headless cloud box — needs re-auth handling.
  3. It's only as smart as the memory you give it — so Half A (memory-perfect) is the prerequisite for it to match your laptop.

🔒 Vulnerabilities / blind spots / security (you asked — this matters most)

A cloud box holding your data + credentials + write-access to your accounts is a real, new attack surface — bigger than your laptop (which is off most of the time and physically yours). Honest risks + the guards:

• The command channel is the door. Your bot @username is PUBLIC (anyone can find it); the ONLY gate is the chat-id allowlist (memory: feedback_bot_username_public_allowlist). The cloud Worker's channel MUST be allowlisted to YOUR chat-id only — otherwise a stranger could command your Worker.
• Credentials on the box. It would hold API keys / MCP tokens / possibly Plaid/bank read access. A breach = serious. Guards: minimal + scoped tokens, secrets in a vault (never plaintext), rotate, no broad bank write access.
• Autonomy blind spot. An always-on agent can act while you're away — things you didn't watch happen. Guard: keep destructive / external / money actions confirmation-gated, log everything, you review.
• The box itself: SSH-keys-only (no passwords), firewall, no public endpoints except one authenticated one, OS/deps patched.
• Honest bottom line: manageable with discipline, but NOT zero-risk — for someone who (rightly) doesn't want to be exposed, this deserves a proper locked-down setup, not a quick spin-up.

THE SEQUENCE (Worker 🐝 recommendation)

Don't rush a box online with your credentials. Order:
1. Finish memory-perfect (Half A) — you want it anyway, and it's what makes the cloud Worker actually smart.
2. Then stand up the cloud Worker (Half B) as a clean, security-locked step (allowlist · scoped secrets · confirmation gates · no plaintext keys).
3. The cloud Worker basically falls out of memory-perfect once the brain is portable.

Decision for Sam: GO now (accept the setup + a short security build), or after the current fires (Asher/cash/MIS report) — which is what I'd advise, since none of today's money problems are solved by the cloud box. Tracked as a card: "Cloud Brain plan."

Related: queue card #042 (memory cross-device sync) · GitHub #24 (OneDrive→Drive) · #17 (self-learning loop, .remember broken) · memory project_worker_b_gabbai.