Matters

Inspiration

People live inside situations, but computers usually store only fragments of them. A travel plan may be scattered across booking emails, ticket PDFs, calendar dates, messages, and a folder of documents. A job search may be spread across a résumé, application records, recruiter emails, interview notes, and follow-up tasks. A long project may exist partly in source files, partly in conversations, and partly in the person's own understanding of what has already happened and what still needs attention.

The person can usually reconstruct the larger picture in their head, although even that takes effort when the material is scattered. An AI has a harder problem: it may be able to read one file or one email, but it does not automatically understand which larger situation that clue belongs to, what stage the situation has reached, or what remains unresolved. The user often has to explain the same background again before the AI can help.

That is the problem I built Matters to address. It is not designed to remember every conversation or preserve every small detail. It is designed to reconstruct the important situations a person is currently inside and give both the person and the AI a clear view of where those situations stand.

What I built

Matters is a local-first human–AI situation workspace. A compatible AI host such as Codex—or a properly configured environment such as Claude Code—examines clues from sources the user has explicitly authorized. It relates those clues and builds a smaller number of meaningful “Matters”: a trip, a job search, a long-running project, a major purchase, a family obligation, or any other situation with its own history, current state, and next steps.

A Matter is not a folder full of copied information. It is a model of the larger situation. It can show the people involved, important events, current stage, child Matters, timeline, relationships, plans, unresolved questions, outcomes, and the sources that support the overview. Small records remain clues, evidence, work items, or events; they do not all become separate Matters and they are not all reproduced in the interface.

For example:

• A travel Matter can bring together the destination, dates, bookings, companions, missing documents, and remaining decisions without trying to store every sentence from every travel email.
• A job-search Matter can show the roles being pursued, application stages, contacts, interviews, results, and follow-ups across otherwise separate files and messages.
• The Build Week Matter in this submission connects the decision to participate, the corrected deadline, the work stages, child Matters, source messages, and the final submission state.

The result is an orientation map. It answers the larger questions—“What is this situation?”, “Where are we now?”, “How did we get here?”, and “What still matters?”—while keeping important conclusions traceable to their supporting clues.

How it works

The first experience begins in conversation with the AI. The user says, “Install and use Matters,” and tells the AI what it is allowed to investigate, such as particular folders, accounts, projects, or connected sources. In the current version, this authorization is given to the AI during installation; it is not entered into a settings screen inside the Windows app.

After that explicit request, the AI completes the technical setup. It installs Matters, connects the public Matters skill and MCP gateway, prepares the private runtime, creates the single daily maintenance task, performs the first modeling run, and opens the desktop interface. The user does not have to install each component separately or create a scheduled task by hand.

During modeling, Matters registers authorized sources in place, filters out material that should not enter semantic analysis, and looks for relationships among the remaining clues. The AI then proposes the larger Matters and their internal structure. This is a selective process: discovering a file does not mean that the file deserves its own Matter, and reading a message does not mean that every detail should become part of the model.

The Windows application is the human entrance to the resulting situation map. It presents a bilingual catalog of Matters and lets the person explore cards, child Matters, timelines, people, sources, images, and AI supplemental context. The AI uses a separate gateway to read the same model, leave observations, submit corrections, compare predictions with later outcomes, and report places where the model is incomplete. There is no second hidden version of the situation that only the AI can see.

The daily task keeps the high-level picture current as authorized sources change. It follows the same investigation boundary the user stated during installation. Merely opening Matters.exe does not authorize an investigation, create a task, or build a model. If the EXE is opened before the AI has completed the first run, an empty catalog is a normal view of an empty private model store.

Why this is different

Search is useful when someone already knows what document they want. A task manager is useful when someone manually enters the work. Matters serves a different need: it reconstructs the situation that connects the documents, messages, events, people, and tasks.

This is also different from asking an AI to remember everything. The goal is not total recall. Matters deliberately compresses many low-level clues into a human-scale model while preserving enough provenance to inspect how the overview was formed. A supported inference about the past remains visibly provisional, a future obligation remains planned rather than completed, and a correction changes the current understanding without erasing its history.

That shared overview helps both sides. The AI can understand the user's current context before acting, rather than asking the user to narrate the entire background again. The person can see the same map, recognize when the AI has grouped something incorrectly, and correct it. Human–AI collaboration becomes less like handing work to a black box and more like maintaining an understandable model of the situation together.

What the demo shows

I used the Build Week submission itself as the demo Matter. The video starts from the main catalog, opens the Build Week Matter, and moves through its overview, child Matters, timeline, AI supplemental information, and supporting details. Rather than displaying every file or every conversation, it shows how those clues were organized into the larger Build Week situation and its current state.

This makes the demonstration intentionally circular: the project being judged is also one of the real situations that Matters helped Codex understand and follow while I was building it.

Privacy and trust

Matters is local-first. Original files and messages stay where the operating system or provider already stores them. Stable references, indexes, derived understanding, and the UI projection live in a separate private MATTERS_HOME. The public repository and release do not contain the user's real sources or private Matter model.

Authorization is narrow and explicit. During installation, the user tells the AI exactly what it may investigate. Permission to read those sources does not grant permission to modify them, delete them, execute anything, send messages, disclose information, or publish it. The initial model and the later daily maintenance task must remain inside the same stated boundary.

How I built it with Codex and GPT-5.6

Codex with GPT-5.6 was involved throughout Build Week. I used it to explore the architecture, define the distinction between a Matter and a smaller clue or event, implement the Python and SQLite core, build the bilingual desktop experience, design the public skill and MCP gateway, generate tests, diagnose regressions, package the Windows application, and prepare the submission.

The most valuable part was not isolated code generation. Codex worked across an evolving repository and an evolving real-world Matter, helping me connect requirements, implementation decisions, evidence, tests, corrections, and release work. Matters is therefore both a product built with Codex and a practical attempt to give Codex a better model of the larger human situation in which its work takes place.

Challenges and lessons

The hardest problem was choosing the right level of abstraction. If every email becomes a Matter, the product is only another inbox. If the model is too broad, it becomes a vague summary. I had to distinguish a major situation from its stages, events, evidence, and small work items, while still preserving the links that let someone audit the result.

I also had to make one model useful to two very different readers. A person needs a calm, legible overview; an AI needs structured, bounded context. Other challenges included separating observation from inference, keeping the English and Chinese views aligned, rejecting unrelated source contamination, and keeping private runtime data out of the public repository and release.

The main lesson was that an AI does not need every detail to understand someone better. It needs the right model of the situations that give those details meaning.

What I am proud of

A few parts of the result feel especially important to me:

• Matters turns scattered clues into a manageable set of high-level situations instead of trying to reproduce a person's entire information history.
• The Windows interface and the AI gateway use the same Matter model, so the overview available to the AI is also visible and correctable by the person.
• Sources, uncertainty, corrections, changing evidence, predictions, outcomes, and model gaps remain traceable without overwhelming the main interface.
• One natural-language installation request lets the AI connect the product, create the single daily task, build the first model within the stated scope, and open the interface.

Together with the bilingual experience, child Matters, timelines, eleven focused maintenance skills, and the public Matters skill for Codex, these pieces make the project feel like the beginning of a genuine situation layer for human–AI collaboration.

What's next

I want to improve how Matters identifies the right high-level situations, make the first-run authorization conversation clearer, add more privacy-preserving connectors, and give users better ways to review or revise the scope they originally gave the AI. I also want to deepen prediction feedback so the model becomes better at showing not only what is happening, but where its earlier understanding was incomplete.

Built With

• codex
• css
• gpt-5.6
• html
• javascript
• model-context-protocol-(mcp)
• python
• sqlite