Private AI Journal 2026: Ollama + Obsidian Setup, Zero Cloud

If you’ve ever told a chatbot something you wouldn’t say out loud, pay attention to this week’s news. A woman’s private Claude conversations were reportedly handed to police, and many people now want to know where their AI chats are stored and who can get to them. If you use AI for journaling, therapy-style reflection or working through a hard decision, ask that question before you type, not after. This guide shows how to build a private AI journal that runs entirely on your laptop with Ollama, Gemma 3 and Obsidian. Nothing you write leaves the machine.

What’s actually new: why a private AI journal suddenly matters

Few details of the reported case have been publicly confirmed, and we won’t guess at them. The general point holds however the specifics turn out. When you chat with a cloud AI service, your conversation sits on someone else’s servers under someone else’s retention policy. As with email, cloud documents and search history, authorities can reach that data through legal process such as subpoenas, warrants and court orders. Most AI providers state in their privacy policies that they comply with valid legal requests. Few users read that far.

The law hasn’t changed. How people use AI has. Two years ago most prompts were things like “rewrite this email.” Now people use chatbots as journals, confidants and late-night sounding boards. A journal entry about a fight with your partner, a health scare or a money problem used to sit in a paper notebook in your drawer. Now it often sits in a cloud database next to millions of other people’s entries. That greatly increases your exposure, even if you have done nothing wrong.

Local AI has also matured enough that you no longer have to accept that trade-off. Open-weight models like Google’s Gemma 3 run well on an ordinary laptop. Ollama installs one with a single command. Obsidian already stores your notes as plain Markdown files on your own disk. Combine the three and you get a reflective AI companion with no account, no server and no logs held by anyone else.

Why it matters

• Data that isn’t collected can’t be handed over. In a fully local setup, no provider holds your journal, so none can respond to a request for it. Your exposure shrinks to your own device, which you control and can encrypt.
• “Delete” means delete. A deleted cloud chat may linger in backups, abuse-monitoring logs or legal holds. A deleted Markdown file on an encrypted drive is actually gone.
• Your entries never become training data. Cloud AI policies change, even with opt-out settings. With offline AI journaling, the question never comes up.
• It works offline. You can journal on a plane, in a cabin or during an outage. Local LLM privacy brings reliability with it.
• It costs nothing to run. No subscription, no per-token billing, no usage caps. Your hardware is the only limit.
• Local still has risks. Someone can still search a seized or stolen laptop. Full-disk encryption and a strong password are core parts of this setup, not optional extras.

How to use it today: Ollama Obsidian setup, step by step

Expect about 20 minutes. You need a Mac (Apple Silicon is ideal), a Windows PC or a Linux machine with at least 8 GB of RAM. Use 16 GB for the stronger 12B model.

1. Turn on full-disk encryption first. A journal on an unencrypted drive isn’t private. On macOS, open System Settings, then Privacy & Security, then FileVault. On Windows, turn on BitLocker or Device Encryption. On Linux, use LUKS (usually chosen at install time). To check on Windows:

   manage-bde -status C:

2. Install Ollama. On macOS or Windows, download the installer from ollama.com, or use a package manager:

   # macOS (Homebrew)
   brew install ollama
   
   # Windows (winget)
   winget install Ollama.Ollama
   
   # Linux
   curl -fsSL https://ollama.com/install.sh | sh

   Then confirm the install:

   ollama --version

3. Pull Gemma 3. Pick a size that fits your RAM. These Gemma 3 Ollama commands download the model weights once. After that, everything runs offline.

   # 8 GB RAM: fast, decent quality
   ollama pull gemma3:4b
   
   # 16 GB+ RAM: noticeably better reflection and nuance
   ollama pull gemma3:12b
   
   # Test it
   ollama run gemma3:4b "In one sentence, what is journaling for?"

4. Lock Ollama to your machine. By default Ollama listens on 127.0.0.1:11434, so only your computer can reach it. Keep it that way. Do not set OLLAMA_HOST=0.0.0.0, which exposes it to your network. Check what it’s bound to:

   # macOS / Linux
   lsof -i :11434
   
   # Windows (PowerShell)
   netstat -ano | findstr 11434

   You want to see 127.0.0.1, not 0.0.0.0. One more warning: newer Ollama versions offer cloud-hosted models, usually tagged -cloud. Those run on remote servers. For a private journal, use only models you have pulled locally. Check with:

   ollama list

5. Create an Obsidian vault in a non-synced folder. Install Obsidian from obsidian.md and create a new vault. Keep it out of iCloud Drive, OneDrive, Dropbox and Google Drive. A synced folder sends your journal back to the cloud and defeats the whole setup. A safe path looks like:

   # macOS / Linux
   ~/PrivateJournal
   
   # Windows
   C:\Users\YOU\PrivateJournal

   Leave Obsidian Sync off for this vault too.

6. Install the Obsidian Local GPT plugin. In Obsidian, go to Settings, then Community plugins, and turn off Restricted mode. Click Browse, search for Local GPT, then install and enable it. Recent versions configure providers through the companion AI Providers plugin. If prompted, install that too, then add an Ollama provider with this URL:

   http://localhost:11434

   Set gemma3:4b or gemma3:12b as the default model. Leave every OpenAI, Anthropic or other remote provider unconfigured.

7. Add custom reflection actions. Local GPT’s settings let you define actions that run on selected text. Paste in these prompts:

   Name: Reflect
   Prompt: You are a calm, non-judgmental journaling companion. Read my entry and ask me three open-ended questions that help me understand what I'm really feeling. Do not give advice unless I ask for it.
   
   Name: Find the pattern
   Prompt: Read this entry and identify any recurring worries, beliefs, or behaviors. Quote my own words back to me. Be direct but kind.
   
   Name: Reframe
   Prompt: Take the most negative thought in this entry and offer two alternative, realistic ways to see the same situation. Keep it under 120 words.
   
   Name: Weekly review
   Prompt: These are my journal entries from the past week. Summarize the main themes, what went well, what drained me, and one small experiment I could try next week.

8. Use a daily template. Turn on Obsidian’s core Daily Notes and Templates plugins, then save this as your template:

   ## {{date}}
   
   **Mood (1-10):**
   **What happened today:**
   
   **What's on my mind:**
   
   **One thing I'm grateful for:**
   
   **AI reflection:**

   Write the entry, select the text, open the command palette with Ctrl/Cmd + P, choose Local GPT, then Reflect. The response lands straight in your note.

9. Verify the zero-cloud claim yourself. Turn off Wi-Fi and run a reflection action. If it works, nothing depends on the internet. For a stricter check, block Obsidian’s outbound traffic with a firewall (Little Snitch or LuLu on macOS, Windows Defender Firewall rules on Windows) and turn off automatic update checks in Obsidian’s settings.

How it compares

Option	Where data lives	Cost	AI quality	Legal exposure
Ollama + Gemma 3 + Obsidian	Your disk only	Free	Good for reflection; weaker at complex reasoning	Only your own device
LM Studio + Obsidian	Your disk only	Free	Similar (same open models)	Only your own device
Claude / ChatGPT / Gemini apps	Provider’s servers	Free to roughly $20+/month	Best available	Provider can be compelled to produce records
Cloud AI journaling apps	App’s servers, often plus a third-party LLM API	Usually subscription	Varies	Two companies may hold your data
Paper notebook	Your drawer	A few dollars	None	Only physical access

Our take: cloud models still clearly beat local ones at hard reasoning, coding and research, so keep using them for that. A journal doesn’t need frontier-level intelligence. It needs good questions and discretion, and a 4B or 12B local model delivers both.

What’s next

Expect AI chat privacy, including police access to AI chats, to become a real policy fight. Privacy advocates argue that conversations with AI deserve protection similar to therapist or attorney communications, and cases like this one add pressure. Until a law says so, assume a court can request anything stored on a provider’s server. Watch for providers adding end-to-end encrypted or zero-retention chat modes, and read the details closely when they do.

Local AI is moving fast. Each new generation of open models gets more capable at the same size, so the gap between a local journal companion and a cloud chatbot keeps shrinking. Ollama keeps adding features, so confirm that each new one, especially anything labeled cloud or sync, still keeps your data local. Obsidian’s plugin ecosystem is also adding local embeddings, which will let you ask “when did I last feel like this?” across years of entries without sending anything off your machine.

Build this habit now: decide deliberately which AI conversations belong in the cloud and which don’t. Send work questions, code and research to the best model available. Keep your inner life on your own encrypted disk.

Frequently Asked Questions

Is a private AI journal with Ollama really 100% offline?

Yes, once you download the model. Ollama runs inference on your own CPU or GPU, and the Local GPT plugin talks only to localhost. To prove it, run a reflection with Wi-Fi off. The two exceptions are Ollama’s optional cloud models and a vault stored in a synced folder. Avoid both.

Can police still get my journal if it’s stored locally?

With a valid warrant, they can seize your device. The difference is that no third party holds a copy it can hand over without your knowledge. Full-disk encryption with a strong password protects you if the device is lost or stolen. This guide isn’t legal advice; talk to a lawyer about your specific situation.

Which Gemma 3 model should I use for journaling?

Use gemma3:4b on machines with 8 GB of RAM and gemma3:12b with 16 GB or more. The 12B model asks noticeably more thoughtful follow-up questions. The 27B model is overkill for journaling unless you have a high-end GPU or a Mac with plenty of unified memory.

Does the Obsidian Local GPT plugin send data anywhere?

It sends requests only to the provider you configure. If your only provider is local Ollama at http://localhost:11434, nothing leaves your machine. Don’t add remote API keys to the journal vault.

Can I use this on my phone?

Not easily. Obsidian runs on mobile, but phones can’t run Ollama the same way. Syncing the vault to your phone usually means a cloud service, which reintroduces the risk. For maximum privacy, keep this journal on your laptop.

Is local LLM privacy worth the quality trade-off?

For journaling, yes. Reflection prompts don’t need frontier reasoning. They need a patient, non-judgmental listener, and Gemma 3 does that well. Save cloud models for tasks that need the extra capability and involve nothing you need to keep private.