Getting Started with Memex: From First Record to First Insight

Memex has more going on under the hood than most journal apps. Multi-agent AI, P.A.R.A. knowledge organization, timeline cards, insight generation — it can sound like a lot when you first open the app. This guide walks you through the first fifteen minutes so you know what to expect and what to do.

Step 1: Install

Memex is available on iOS (App Store) and Android (Google Play). If you want the latest build before it hits the stores, Android APKs are also on GitHub Releases.

The app is about 80MB. No account is required. When you first open it, you will see an empty timeline — that is normal. The app does not pre-populate with sample data.

Step 2: Connect a model provider

This is the step that trips up most new users, so let me be specific about what to do.

Tap the avatar icon in the top-left corner of the timeline, then tap Model Configuration. You will see a provider dropdown. Pick one:

• Google Gemini — easiest to start with. You can use OAuth (sign in with your Google account, no API key needed) or enter a Gemini API key from Google AI Studio. The free tier is generous.
• OpenAI — if you already have an OpenAI account. Enter your API key from platform.openai.com. Also supports OAuth sign-in.
• Anthropic Claude — strong reasoning quality. Get a key from console.anthropic.com.
• Ollama — fully offline, zero cost. Requires Ollama running on a local machine that your phone can reach. Best for technical users.

After selecting a provider, enter your API key (or sign in via OAuth), and optionally adjust the base URL if you are using a custom endpoint. Tap save. The app will verify the connection.

One detail worth knowing: each built-in agent can use a different model. You can configure this later in the agent settings. For now, the default model applies to all agents, which is fine for getting started.

Step 3: Record something

Go back to the timeline and tap the input area at the bottom. You have three ways to record:

• Text — type anything. A thought, a task, an observation. Length does not matter.
• Photo — tap the camera or gallery icon. Memex automatically extracts EXIF data (timestamp, GPS location) and runs on-device OCR and image labeling via Google ML Kit.
• Voice — long-press the mic button to start recording, release to send. Transcription runs entirely on-device using sherpa-onnx. The first time you use voice, the app downloads a speech model (~230MB).

Try recording two or three things. A text note about what you did today, a photo of something nearby, and a quick voice memo. This gives the AI enough variety to demonstrate different card types.

Step 4: Watch the agents work

After you submit a record, the multi-agent pipeline kicks in. You can see the activity in real time — there is an agent activity indicator that shows which agent is currently processing.

Here is what happens behind the scenes:

• The Card Agent analyzes your input and generates the most fitting timeline card type — task, event, place, person, metric, gallery, article, and over fifteen other types.
• The PKM Agent extracts knowledge and files it into your P.A.R.A. knowledge base (Projects, Areas, Resources, Archives) as interconnected Markdown files.
• The Comment Agent adds contextual AI commentary to the card.
• The Insight Agent looks for patterns across your records — this one becomes more useful as you accumulate more data over days and weeks.

The first card usually appears within a few seconds. The knowledge filing takes a bit longer. You do not need to wait — you can keep recording while the agents process in the background.

Step 5: Explore what was generated

After the agents finish, tap on a card in your timeline to see the detail view. You will see the structured card with its type, extracted entities, tags, and any linked knowledge.

Check the other tabs:

• Insights tab — this is where cross-record patterns appear. It will be sparse at first but grows richer as you record more. Insights show up as trend charts, radar maps, timelines, highlight quotes, and narrative summaries.
• Knowledge tab — this shows your P.A.R.A. knowledge base. Browse the directory structure to see how the PKM Agent organized your records into Projects, Areas, Resources, and Archives.

What to expect in the first week

Memex gets more useful over time. Here is a realistic timeline:

• Day 1 — you see individual cards generated from your records. The knowledge base starts forming. Insights are minimal.
• Days 2-3 — the PKM Agent starts linking records across categories. You begin to see connections between entries you did not explicitly create.
• Days 4-7 — the Insight Agent has enough data to surface patterns. You might see a trend chart about a recurring topic, a timeline of places you visited, or a narrative summary of your week.
• Week 2+ — the system becomes genuinely useful for reflection. Cross-record insights get richer. The knowledge base becomes a living document of your life.

The key is consistency. Record a few things each day — it does not need to be long or polished. A sentence, a photo, a voice note. The agents do the rest.

Common first-time questions

The card type seems wrong. It happens. The Card Agent makes its best guess based on content. As models improve, accuracy improves too. You can also try a different model provider if one consistently misclassifies your records.

Voice transcription is slow the first time. The speech model needs to download once (~230MB). After that, transcription runs locally and is fast. If you are in China, there is a mirror option for faster downloads.

I want to change the model later. Go to avatar → Model Configuration anytime. You can also configure different models for different agents — for example, a cheaper model for card generation and a stronger model for insights.

Where is my data stored? Everything is on your device. Markdown files and a local SQLite database. You can back up to iCloud Drive, a custom folder, or keep it in app storage. One-tap backup and restore.

Next steps

Once you are comfortable with the basics:

• Read our LLM selection guide to optimize which model you use for each agent.
• Explore the custom agent system to build your own AI agents on your phone.
• Check the changelog for the latest features and improvements.

The source code is on GitHub. If you run into issues, the Discord community is the fastest way to get help.

FAQ