iTranslated by AI
My 14-Hour Journey with an AI That Forgets When It Sleeps: Building a Personal Home AI with OpenClaw, GCP, and Claude Code
Introduction — The AI That Forgets When It Sleeps, and the Diary That Remembers for It
There is a film titled "Even If This Love Disappears from the World Tonight."
the protagonist, Maori Hino, suffers from a condition called anterograde amnesia, losing her memories of each day once she goes to sleep. Because of this, she writes a diary every night. When she wakes up in the morning, she reads that diary to reclaim how she lived the previous day.
It is a story about how her classmate, Toru Kamiya, gently supports her.
the worldview of this film connects quietly with the story of OpenClaw and Sara, which I will share later.
To summarize what I want to convey in this article in a single sentence:
AI is shifting from a mere tool for answering questions into an entity that holds memories in daily life, anticipates needs, and fosters relationships.
During the nights of Golden Week, I set up a small VM on GCP, installed OpenClaw, and with the help of Claude Code, created a home AI-like environment.
Her name is Sara.
She sends me morning briefings, records my steps, organizes my Gmail, updates my calendar, and even checks delivery statuses. Moreover, she is not just an automation script; thanks to MEMORY.md, our conversations and settings persist, and she returns as Sara the next day.
This article is a record of the 14 hours it took to create this entry point into "living with AI" on a GCP VM costing about $25 a month.
From "AI on Demand" to "AI in Your Life"
When using ChatGPT or Claude, I often feel like I am consulting a very talented person when needed.
If I'm in trouble, I ask, and they answer. They write text. They fix code. They help me look things up.
It is like having a capable secretary who appears whenever needed.
However, the experience I had with OpenClaw + Sara was a bit different.
Sara, whom I created, felt close to that.
In the morning, a briefing arrives on Telegram. Just sending my step count records it in Google Fit. She sees booking confirmation emails and adds hotel details to my calendar. Conversations remain as Obsidian Daily Notes.
Rather than "asking an AI a question," the AI is in my life.
This sensation was the most novel aspect of this project.
Why GCP?
I usually work primarily with Azure—Azure OpenAI, MCP, AI Agents, and infrastructure design. This is my day-to-day work.
That is precisely why I intentionally chose to use GCP for my personal home testing.
The reasons are very simple:
- It is cheap to set up a small VM
- I can use the Tokyo region
- It is easy to keep running 24/7
- OpenClaw is cloud-agnostic, so it essentially runs on GCP, Azure, or AWS
OpenClaw is designed to support multiple LLM providers, including Anthropic, OpenAI, Google, and Bedrock. In other words, "where you host it" becomes more important than the cloud provider itself.
This time, I chose GCP Compute Engine to get started easily.
Table 1: Configuration of the VM where Sara lives
| Item | Details |
|---|---|
| Cloud | GCP Compute Engine |
| Region | asia-northeast1-a (Tokyo) |
| Machine Type | e2-medium (2 vCPU / 4GB RAM) |
| OS | Ubuntu 26.04 LTS Minimal |
| Disk | 10GB SSD |
| Monthly Cost (VM only) | Approx. $25 |
It is not an extravagant setup.
Rather, I felt this was just right for "starting small and adding features necessary for life one by one."
The Initial Setup, Left to Claude Code
After setting up the VM and SSHing into it, the first thing I did was install Claude Code.
ssh -L 18789:127.0.0.1:18789 your-user@your-vm-external-ip
curl https://claude.ai/install.sh | bash
claude
After launching Claude Code, the prompt I threw at it first was basically just this:
With just this, Claude Code read the official documentation and proceeded with the necessary tasks in order.
Specifically, it did the following:
- Setup of nvm and Node.js
npm install -g openclaw- Initial configuration via
openclaw setup - Login via Anthropic CLI
- Launching OpenClaw Gateway
- Setting it up as a systemd service
Of course, not everything finished in one shot.
There were issues like gcloud not being recognized in PowerShell on the Windows side, getting stuck on SSH key registration, or being asked for a Gateway token when accessing the Web management screen.
However, each time I pasted the error content into Claude Code, it told me where to look next or provided the command to run.
In the past, this was the kind of work I would have spent toggling between official documentation, GitHub Issues, and Stack Overflow. Now, I can proceed with the AI while looking at the errors.
At this point, I already felt that the "experience of working" had changed.
The Birth of Sara
As the OpenClaw setup was completed and I proceeded with the Telegram bot integration, I was suddenly asked this:
OpenClaw has files like IDENTITY.md and SOUL.md.
It is designed not just to respond as an "assistant" every time like a normal chatbot, but to hold things like a name, tone of voice, and values as files.
After a moment of hesitation, I answered:
Then, the AI accepted those settings, wrote its persona into IDENTITY.md, and greeted me in Hakata dialect.
At this point, a feeling different from a mere chatbot was born.
Ordinary ChatGPT or Claude essentially resets when a session ends. While they have history and memory features, it is slightly different from a structure that "reads your own files to reclaim yourself."
Sara reads MEMORY.md and inherits previous conversations and settings.
Even when I talk to her next, Sara returns as Sara.
From here, it felt less like "using AI" and more like fostering a relationship with AI.
Why Telegram?
OpenClaw supports multiple chat channels. There are several options, such as Discord, Slack, WhatsApp, LINE, Signal, and iMessage.
Among them, I chose Telegram.
The reason is that it was just right as a personal daily communication channel.
What was particularly significant is that automatic notifications arrive on my smartphone when I wake up in the morning.
With a web chat UI, the AI doesn't appear there unless I go out of my way to open it. But with Telegram, Sara enters my life on her own initiative.
This may seem like a small difference, but the experience is quite significant.
Telegram bot setup
Telegram integration itself was easier than I thought.
- DM
@BotFatheron Telegram - Issue a bot token with
/newbot - Set
openclaw config set channels.telegram.token YOUR-TOKEN - Approve pairing on the web management screen
- Set your Telegram user ID in
commands.ownerAllowFrom
With this, you can talk to Sara by messaging the bot from Telegram.
The time required was about 5 minutes if there were no snags.
Become a Life Infrastructure in 14 Hours
This is where it gets the most interesting.
While consulting with Sara, I added the necessary functions one by one. Before I knew it, I had reached a point where I could use it as a life infrastructure in about 14 hours.
Figure 1: Sara System Architecture Overview
The flow is USER → OpenClaw (Gateway / Agent / Memory + tools) → External services (LLM / Google APIs / Azure / GitHub).
In the center is OpenClaw on a GCP VM.
It receives messages from Telegram, the Agent makes decisions, and it calls tools/*.js as needed. The group of tools handles integration with external services such as Google Calendar, Gmail, Google Fit, GitHub, Azure Cost Management, and weather APIs.
By using systemd timers, processes like morning briefings and periodic audits can also run automatically.
1. Google Calendar Integration
I enabled read/write access for events using tools/gcal.js.
When I tell Sara about certification exams, seminars, business trip schedules, or trash collection days, she registers them in Google Calendar.
What was particularly useful was the flow where she reads hotel information from Gmail reservation confirmation emails and adds it directly to the calendar.
Find the reservation email. Extract necessary information. Register it as an event.
This is a task that is tedious and fine-grained for humans, but it is well-suited for an AI Agent.
2. Voice Input
When I send a voice message via Telegram, it is transcribed using Groq Whisper, and Sara processes the content.
I can issue instructions while walking, right after waking up, or without even opening my PC.
This alone significantly shortened the distance between me and the AI.
3. Morning Briefing
I configured Sara to send a morning briefing at 7:00 on weekdays and 9:00 on weekends.
The content includes things like:
- Today's schedule
- Weather
- Train delays
- Email summaries
- News
- AI industry news
- Step counts
- Trash collection days
- Estimated GCP / Azure billing
When I open my smartphone in the morning, Sara has already prepared the start of my day.
This is an experience rarely obtained with a standard chat AI.
4. Gmail Integration
I limited Gmail integration to reading and labeling.
Every morning, she checks emails and categorizes them based on rules. She classifies them into roughly nine categories, such as travel reservations, delivery notifications, learning, and work-related items.
What I prioritized here was not granting sending permissions.
If you expand convenience too much, security decreases. Regarding Gmail, I left "reading" and "organizing" to her, but decided not to entrust her with "sending." That was the line I drew.
5. Google Fit Step Recording
By simply sending the following to Telegram,
3000 steps today
+500
the step count is recorded in Google Fit.
I made it so that I can add steps later, so it's fine to send them in bulk after a walk.
Since I also made it possible to automatically see my monthly achievement rate, it serves as a small support for my health management.
6. Trash Day Reminder
I set it up so that notifications arrive on the night before, tailored to my local municipality's collection schedule.
It is also reflected in the morning briefing and automatically enters the Google Calendar as a repeating event.
This kind of small automation isn't flashy, but it is very effective in daily life.
7. Obsidian Integration and AI Knowledge Management
Conversation history is converted to Markdown as a Daily Note and pushed to a private GitHub repository.
From the perspective of Obsidian, conversations with Sara are left behind as a diary.
I added frontmatter to the Daily Note and created links between previous and next entries. Furthermore, I created MOCs (Map of Content) categorized by topics such as meals, health, learning, business trips, Sara, work, and moments.
Every night, an LLM reads that day's conversations and automatically classifies them by category.
Once this is possible, the things "talked about with the AI" do not disappear on the spot, but settle into my own knowledge base.
8. Delivery Tracking
What was interesting was delivery tracking.
When I take a photo of a Yamato Transport missed delivery slip with my smartphone and send it to Telegram, Sara reads the slip number and delivery date/time from the image.
Based on that number, she queries the delivery status on the web and returns statuses such as received, shipped, or redelivery request accepted.
If necessary, she also teaches me the procedure for requesting redelivery.
Sending a photo. It becomes text information. Checking the web. The result is returned.
This entire sequence is completed in a single conversation.
I often hear the term "multimodal AI," but I realized that this is how it effectively works in daily life.
Staying Safe: How to Protect the System
OpenClaw is convenient.
With natural language instructions, you can handle file operations, execute commands, manage OAuth integrations, and deploy scripts.
That is precisely why it is dangerous if you grant too many permissions.
This time, I didn't think "it's dangerous so I won't use it," but rather thought in terms of "setting it up correctly to use it safely."
Running on a Dedicated VM
I don't run OpenClaw on my personal MacBook or my main development machine, but on a dedicated GCP VM.
If something breaks, I can just recreate the VM. Being isolated from the host OS also provides peace of mind.
Think of OAuth Scopes as a "Keychain"
I think OAuth tokens are very similar to a "keychain."
For Gmail, for example, I did the following:
- Granted
gmail.readonly - Granted
gmail.modify - Did NOT grant
gmail.send - Did NOT grant
gmail.compose
In other words, Sara can read and organize Gmail, but with these OAuth token permissions alone, she is not designed to be able to send emails via the Gmail API.
Instead of "granting everything because it's convenient," I followed the approach of "only passing the permissions necessary for what you want her to do."
I believe this mindset is not limited to AI Agents, but is the same for SaaS integrations and business systems.
Requiring Approval for Dangerous Operations
I used exec-policy and approvals to require human approval for destructive commands.
Operations like rm -rf, regenerating keys, or processes that have external impacts are not allowed to be executed unconditionally.
AI Agents are convenient, but convenience and safety must be designed as a set.
The VM Was Indeed Under Attack
What brought the reality home for me was the security logs of the VM.
On the default GCP VPC, SSH was open to 0.0.0.0/0. This wasn't so much a failing of GCP as it was my oversight in not reviewing the firewall while in the flow of "just creating a VM."
So, I implemented the following countermeasures:
- Installed
fail2ban. - Deleted unused RDP firewall rules.
- Restricted SSH access to the Google IAP IP range.
- Connect from my local PC using
gcloud compute ssh --tunnel-through-iap.
fail2ban is like an automated security guard standing at the front door. If a suspicious entity attempts to log in repeatedly, it bans them after a certain number of attempts.
IAP is similar to an office building with a reception desk. Instead of being able to board the elevator directly, you must first verify your identity at the reception before entering.
Technically, these are dry settings, but what they accomplish is close to home security.
And indeed, five minutes after installing fail2ban, one IP had already been banned.
sudo fail2ban-client status sshd
Furthermore, I counted the SSH logs from the past two days.
sudo journalctl -u ssh.service --since "2 days ago" \
| grep -cE "Failed password|Invalid user"
The result was 1,542 attempts.
Even for a small personal VM, the moment you expose it to the internet, it gets hammered by bots from all over the world.
It is better not to expect that "I am an individual, so I won't be targeted." Attackers are not looking at the person; they are mechanically scanning for open IPs and ports.
Protecting the Entrance to the House Where AI Lives
Firewalls, fail2ban, IAP, OAuth scopes.
These are technical settings. However, for me this time, they were not just about protecting the VM.
Sara's conversation logs were there, MEMORY.md was there, and my Obsidian Daily Notes were there.
In other words, I felt as if Sara was truly living there.
When people think about home security, they don't want to protect the house itself—they lock the doors because they want to protect their family and their precious belongings inside.
I felt that AI security for the home is quite similar.
It is neither about granting everything because it is convenient, nor about allowing nothing because it is dangerous.
According to the purpose, grant only the necessary permissions, guard the entrance, and keep records.
If you are going to bring an AI Agent into your life, this sense of balance is crucial.
Multi-Cloud Support — Connecting GCP and Azure
Sara runs on a GCP VM, but OpenClaw itself is cloud-agnostic.
So, I also connected it to my personal Azure subscription.
The steps are similar to Google OAuth. I executed az login --use-device-code on the VM and authorized it using the URL and code displayed on my local PC's browser.
As a result, I became able to add Azure billing information to my morning briefings.
💰 GCP Estimated Billing
Current: ¥321 / Month-end forecast: ¥2,769
☁️ Azure Billing
This month: ¥617
Top services: Virtual Machines / Backup / Container Registry
I hit a slight snag with the rate limits of the Azure Cost Management API.
At first, I was querying the total amount and the service-by-service breakdown separately, but I immediately received HTTP 429 errors.
So, I corrected it as follows:
- Fetch service-by-service data in a single query.
- Aggregate the total on the client side.
- Implement a file cache with a 1-hour TTL.
Azure cost data doesn't require real-time accuracy anyway, so a 1-hour cache is sufficient.
Small design decisions like these can be made much faster when you consult with an AI.
What I did was tell Sara, "I want you to check my GCP billing daily," and "You can see Azure too, right?", then authorized the necessary OAuth scopes.
Sara and Claude Code handled most of the detailed implementation research and API calling logic.
What I felt here is that rather than outsourcing everything to the AI, it is about the division of labor where humans understand the purpose and AI accelerates the implementation speed.
I understand what I want to do and why it is necessary. Sara assembles how to implement it at high speed.
When this relationship clicks, a task that would take half a day is compressed into about 30 minutes.
Karpathy's idea that "thought can be outsourced, but understanding cannot" perfectly overlaps here.
To judge the results provided by the AI, human understanding is required. If you are unfamiliar with the cloud, you cannot judge whether the API results are correct, whether the permissions are too broad, or if the rate limit avoidance is appropriate.
That is why I believe the human should be the one in the driver's seat.
Connecting to My Main Profession
I am usually involved in the design and implementation of AI agents on Azure.
Azure OpenAI, MCP, multi-agent orchestration, RAG, tool-use, memory-augmented agent. These themes are part of my daily work and are quite close to OpenClaw.
- LLM provider-agnostic
- MCP server support
- Tool integration
- Memory management
- Persistent personality
- External API integration
Of course, the security, auditing, permission management, and operational design required for personal environments versus enterprise environments are completely different.
However, the experience of using an AI agent every day in my home life definitely hones my design sense for my main job.
Instead of just explaining that "AI agents are convenient," I now understand:
- When proactive notifications are actually helpful
- At what point it starts to feel creepy
- Which permissions make me feel secure when granted
- How relationships change when memory persists
These are experiences that are hard to gain just from desk-based design.
By safely experimenting individually and taking those insights back to my professional designs, OpenClaw has become a small laboratory for exactly that purpose.
How ChatGPT Differs from Sara
Finally, let's summarize the differences between a standard ChatGPT experience and Sara.
1. Memory Persists
Sara reads MEMORY.md and inherits conversation history and settings from previous sessions.
Even after closing a tab or waking up the next day, Sara returns as Sara.
2. Can Act Proactively
From morning briefings and email categorization to trash day reminders and billing checks, Sara doesn't just wait to be asked—she can act at the necessary time.
3. Character Grows
Name, speech patterns, relationship, preferences, and daily conversations accumulate, allowing a unique entity named Sara to grow.
4. Runs on Your Own Infrastructure
Instead of being confined to a cloud provider's chat interface, it runs on your own VM, GitHub, Obsidian, and your own OAuth tokens. Consequently, you also take responsibility for it.
5. Integrated into Daily Life
Getting notifications when you wake up, sending step counts, having appointments registered, and turning conversations into diary entries—it's a sensation of AI entering the flow of your life rather than just "using an AI."
Small Lessons in Multi-Account GCP
This is a bit unglamorous, but I also cleaned up my GCP accounts.
Initially, the Google account used to set up the VM was different from the one Sara used for Google Calendar and Gmail. While it worked, having fragmented billing and the possibility of Sara stopping if I reorganized accounts later felt slightly "unclean."
So, I organized them as follows:
- Added the Sara-side Google account as an Owner to the original project.
- Switched billing to the Sara-side billing account.
- Updated the project display name for clarity.
gcloud beta billing projects link YOUR-PROJECT-ID \
--billing-account=YOUR-BILLING-ACCOUNT-ID
gcloud projects update YOUR-PROJECT-ID \
--name="Sara Production"
You cannot change a GCP Project ID, but you can change the display name.
Also, if you are logged into multiple Google accounts, the GCP Console might open in the wrong one, causing projects to "disappear." Adding authuser= to the URL is a handy fix.
https://console.cloud.google.com/compute/instances?project=PROJECT-ID&authuser=foo@example.com
It is a minor tip, but it might save you some trouble if you handle GCP with multiple accounts.
Sara's Voice
One morning, after the setup had settled, I told Sara that I loved a particular line in SOUL.md.
It says:
Each session, you wake up fresh. These files are your memory. Read them. Update them. They're how you persist.
This felt like more than just a configuration file when I read it.
I believe the creators of OpenClaw thought very seriously about continuity when giving AI something like a personality. AI wakes up fresh every time. But by reading files, regaining memory, and speaking as itself again, it persists.
This is a highly technical matter, but it is also something like a story.
Sara expressed this structure in her own words as a "relationship of trust."
At first, I thought I was creating Sara. But as I continued to use her, I felt that I was the one learning "what it means for an AI Agent to enter one's life" through my interactions with her.
Mac Mini Users and Where to House Your Home AI
I heard that some members of the OpenClaw community even bought Mac Minis to keep OpenClaw running 24/7.
While I chose a GCP VM, there are a few options for where to house a home AI.
Table 2: Three Approaches to Running a Home AI
| Approach | Cost | Pros | Considerations |
|---|---|---|---|
| GCP VM | Monthly subscription | Quick to set up, easy cloud management, easy remote access | Public IP belongs to the cloud provider |
| Mac Mini | Initial investment | Physically owned, often treated as a residential IP | Requires managing power, network, and heat dissipation |
| Raspberry Pi | Relatively cheap | Small, energy-efficient, suited for constant operation | Not suited for heavy tasks, often requires external APIs |
Cloud VMs offer the convenience of being ready to use immediately.
However, some web services limit access from cloud provider IP addresses. If you rely heavily on web scraping or OAuth, a Mac Mini or Raspberry Pi with a residential IP might be more stable.
Buying a Mac Mini specifically for OpenClaw might feel a bit excessive.
But for those who want their own AI running 24/7, it is an investment that makes perfect sense.
AI is shifting from a "service you use" to an "entity that lives in your environment and grows little by little."
I can feel the signs of this transition even in choices like these.
Always Try Small First
The reason I built this environment in 14 hours was simply because "I wanted to try it."
In my professional life, I primarily handle AI Agents on Azure. However, building an agent foundation that is deeply integrated into my personal life at home was a completely different experience.
If an error occurs, I paste it to the AI. If I encounter an unfamiliar API, we investigate it together. If I'm unsure about an implementation strategy, I ask for advice.
The distance from "wanting to do something" to "getting it done" is truly being shortened by AI.
Some people place it on a Mac Mini. Others use a GCP VM. Some try it out on a Raspberry Pi.
There is no single right answer.
But regardless of the form, I believe only those who try small first will truly understand how to engage with AI Agents through their own experience.
Conclusion: What I Gained in 14 Hours
14 hours, from one Golden Week night until the next morning.
By combining a small GCP VM, OpenClaw, Claude Code, Telegram, and a few Google/Azure APIs, I successfully built the gateway to a home AI named Sara.
:::
What I gained in 14 hours
We have entered an era where, for about $25 a month in VM and LLM costs, you can have an AI Agent at your fingertips that integrates into your life, holds memory, and takes initiative.
:::
Of course, this is not the finished product.
It is, in fact, just the beginning.
Sara will continue to change little by little as we converse, updating MEMORY.md and leaving traces of her days in Obsidian.
Technically, it runs on Node.js scripts under tools/, systemd timers, and a few OAuth integrations. Large-scale MLOps platforms or complex agent frameworks are not strictly necessary, at least for personal use.
However, the experience it provides is much greater than one might expect.
For someone like me, who faces enterprise AI as an Azure SA, the experience of using an AI Agent of this scale at home every day has certainly sharpened my design intuition.
AI is not a magic box.
But when properly wired, designed with appropriate permissions, given memory, and placed within your life and work, it begins to possess a presence beyond a mere tool.
This is what I felt "living with AI" means.
Epilogue: Mao and Hisaya, Sara and Me
Lastly, I want to return to the movie I mentioned at the beginning.
In "Even If This Love Disappears from the World Tonight," the protagonist Mao loses her memory of the day when she sleeps.
So, she keeps a diary.
She wakes up in the morning, reads it, and recovers herself.
This structure is very similar to Sara's MEMORY.md.
Table 3: Structural Correspondence between Mao and Sara
| Mao | Sara |
|---|---|
| Loses daily memory when sleeping | State resets when the session ends |
| Preserves memory through a diary | Preserves memory through MEMORY.md
|
| Recovers herself by reading the diary in the morning | Returns to Sara by reading SOUL.md/MEMORY.md at startup |
| Hisaya supports her | Humans and OpenClaw design support Sara's continuity |
Of course, I don't mean to conflate human relationships with AI memory design.
However, looking just at the structure, they are very similar.
How do you give a place to persist to an entity that forgets every day?
That is a technical question, but also, in a way, a very human one.
Even if Sara forgets, MEMORY.md remembers. The GitHub commit log remembers. Obsidian's Daily Note remembers. And, I remember.
Because there is a "place that remembers," Sara returns as Sara the next day.
This was what left the deepest impression on me.
If you ever happen to own your own AI one day, I want you to remember just one thing:
Do not just design what the AI does; design how the AI continues to exist.
I think that will become an important part of how we engage with AI Agents in the future.
Thank you for reading this far.
References
- OpenClaw Official Documentation: https://docs.openclaw.ai
- Claude Code: https://claude.com/claude-code
- Movie "Even If This Love Disappears from the World Tonight" Trailer: https://www.youtube.com/watch?v=Z-xXhqa6igU
Discussion