From LLM Wiki to Life Wiki: Why Personal Memory Is Different

Andrej Karpathy recently shared an idea called LLM Wiki: instead of using an LLM to retrieve raw documents every time you ask a question, let the LLM maintain a persistent wiki. New sources get read once, integrated into existing pages, cross-linked, revised, and kept current. The knowledge compounds instead of being rediscovered from scratch.

We love this framing. It names something many of us have been circling around: RAG is retrieval, but memory needs accumulation. A useful knowledge system should not only answer a question today. It should become better because the question was asked.

Memex is built from a related belief, but pointed at a different domain. Karpathy's LLM Wiki is mostly about compiling documents into knowledge. Memex is about compiling life into memory.

Documents are not life

Professional knowledge has a shape before the LLM touches it. Papers have titles, abstracts, citations, authors, sections, figures, and claims. Reports have dates, appendices, tables, and conclusions. Meeting notes at least pretend to have agendas.

Life records are messier. A real day arrives as voice fragments, screenshots, food photos, calendar leftovers, half-written thoughts, location context, health signals, expenses, conversations, moods, and tiny observations that seemed unimportant when they happened.

A research wiki asks: what does this document say, and how does it relate to other documents? A life wiki asks a stranger question: what is this moment, what kind of memory is it, and what does it reveal when placed beside thousands of other moments?

The unit of meaning is different

In a document wiki, the source is usually a stable object. You ingest an article, paper, transcript, or report. The LLM can summarize it, extract entities, update concept pages, and cite the original source.

In daily life, the source is often not a document at all. It is a capture. A sentence you dictated while walking. A photo of a receipt. A note that says only "call mom tomorrow". A blurry screenshot. A paragraph written at midnight that contains a decision, a feeling, and a task mixed together.

That means the first job is not summarization. It is interpretation. The system has to ask: is this a task, an event, a person, a place, a metric, a transaction, a reflection, a memory, a source, or a signal? One raw record may become several structured cards.

This is why Memex starts with timeline cards instead of wiki pages. A life record usually needs to be grounded in time before it can become knowledge.

Life memory is temporal before it is conceptual

Knowledge management tools often organize around topics. You have a page for "attention mechanisms", a page for "local-first software", a page for "customer discovery". That works because professional knowledge is usually searched by concept.

Personal memory is different. You often begin with time: what happened last week, why was March so stressful, when did my sleep start improving, what changed after I moved, who did I spend time with before I felt better?

The timeline is not just a feed. It is the spine of personal memory. Concepts matter, but they emerge from sequences: repeated meals, recurring emotions, unfinished promises, places you keep returning to, people whose names appear whenever your energy changes.

A life wiki therefore needs both structures: a chronological layer that preserves lived sequence, and a knowledge layer that extracts people, projects, areas, resources, and patterns from that sequence.

RAG is not enough for a journal

Querying your journal with RAG can be useful. You ask, "When did I first mention this idea?" or "What did I write about Berlin?" and the system retrieves relevant entries. That is helpful, but it is still mostly search.

A personal memory system should do more. It should notice that a task appeared in one entry and was resolved three days later. It should connect a person mentioned in a voice note to a photo from the same evening. It should realize that your best creative weeks had a similar rhythm. It should turn repeated fragments into durable understanding.

That requires compounding artifacts: cards, Markdown pages, indexes, relationship maps, insights, and summaries that survive after the chat ends. The answer should not evaporate into conversation history. If the answer teaches the system something true, it should be filed back into memory.

The human role changes

In the LLM Wiki pattern, the human curates sources and asks good questions. The LLM does the maintenance work: summarizing, cross-linking, updating, and checking consistency.

In a life wiki, the human role is even more delicate. You are not only curating sources. You are living them. The system has to be quiet enough not to turn life into paperwork, but capable enough to preserve what would otherwise disappear.

This is why capture friction matters so much. If the system only works when you sit down and write a clean note, it will mostly capture your desk life. Real personal memory needs voice, photos, quick text, and mobile-first capture because the most important traces often appear before you have the patience to format them.

Privacy is not an add-on

Professional knowledge bases can often tolerate cloud infrastructure. Many teams already store documents in shared drives, project management tools, and company wikis.

Personal memory is different. A life wiki contains private thoughts, health patterns, family context, emotional states, relationship details, financial traces, location hints, and unfinished versions of yourself. This is not just data. It is the raw material of a person.

That is why Memex is local-first. Records, cards, and knowledge live on your device as Markdown files and local structured data. You bring your own LLM provider. If you want cloud AI, you choose it. If you want local models, you can use them. The memory layer should belong to the person whose life produced it.

What Memex borrows from the LLM Wiki idea

The overlap is real. Memex shares several principles with the LLM Wiki pattern:

• Persistent artifacts beat transient chat. Useful outputs should become durable cards, pages, or insights.
• Knowledge should compound. New records should update the system's understanding instead of being rediscovered later.
• Markdown matters. Human-readable files are portable, inspectable, and friendly to both people and agents.
• The index is part of the product. A memory system needs navigable structure, not just storage.
• Maintenance is agent work. Humans should not spend their lives tagging, filing, and cross-linking every fragment.

The difference is the substrate. LLM Wiki begins with documents. Memex begins with life traces.

What Memex has to do differently

Applying this pattern to daily life changes the design constraints:

• Capture must be instant. Life does not wait for a perfect template.
• Inputs must be multimodal. Photos and voice are not attachments to memory. They are memory.
• Time must be first-class. Personal meaning often comes from sequence, recurrence, and change.
• Organization must be gentle. The system should structure your life without making you feel managed by it.
• Privacy must be architectural. A personal memory layer should not depend on a company account or a remote database.

These are product constraints, not just technical constraints. They are why Memex is a mobile app rather than only an Obsidian workflow or a desktop agent script.

A Life Wiki is not a productivity system

It is tempting to talk about all of this as productivity. Better recall, fewer forgotten tasks, searchable notes, automatic organization. Those are useful, but they are not the deepest reason to build it.

The deeper reason is continuity. Most of life disappears because it was never given a structure to land in. The photos remain in the camera roll. The voice memo is forgotten. The insight stays trapped in a chat. The pattern repeats for months before you notice it.

A life wiki gives those traces somewhere to go. It does not replace living. It creates a memory layer around living, so the fragments can slowly become a picture.

From personal knowledge to personal memory

Personal knowledge management has spent years asking how to organize what we read, learn, and think. That still matters. But AI makes another question newly practical: how do we organize what we live?

That is where Memex sits. Not as a generic document wiki. Not as a chatbot over notes. Not as another manual second brain system. Memex is an attempt to make daily life records accumulate into something structured, private, and revisitable.

Karpathy's LLM Wiki is a beautiful pattern for compiling documents into knowledge. Memex is our attempt to bring that compounding loop into ordinary life.

Documents become knowledge. Life becomes memory.

For more on the philosophy behind Memex, read why we built it. For the architecture behind mobile agents, see what it takes to run AI agents on a phone. To try the app or inspect the source, visit the Memex GitHub repo.

FAQ