Qorx Zero

Keep the context. Send the proof.

AI coding tools are good at the next task, but serious projects carry decisions across many tasks: what ships first, which constraints cannot move, what was already tried, and what must be forgotten. Re-sending everything wastes context. Sending everything to a remote memory service gives up control.

Qorx Zero gives project memory one clear home: the developer's device.

What it does

Qorx Zero stores project decisions in IndexedDB and makes every retrieval decision visible. Ask a question and the proof desk shows:

• which records matched
• the score and matched terms
• each source hash
• the exact character budget
• the complete proof frame allowed to leave the device

A record can expire automatically or disappear immediately with one tap. Ask for something the local memory cannot support and Qorx returns an empty frame. No proof means no claim.

How it works

A deterministic ranker combines keyword relevance, importance, and recency. Expired records are deleted before ranking. Only records sharing a meaningful term with the question are eligible. The final frame is capped at five records and 1,600 characters.

The full memory store remains in the browser. The server receives only the current question and that visible frame. Its OpenAI Responses API adapter uses GPT-5.6 Terra with low reasoning, low verbosity, and storage disabled. GPT-5.6 is instructed to answer only from the supplied proof and cite the visible source hashes.

Built with Codex and GPT-5.6

Codex with GPT-5.6 helped define the clean-room boundary, implement the IndexedDB store and recall algorithm, write seven automated tests and an independent validation notebook, review the provider request against OpenAI documentation, and prepare the public demo and submission package.

The decisive product choices stayed human: memory must remain inspectable, forgetting must be immediate, missing evidence must produce a refusal, and the public repository must stand alone without exposing proprietary Qorx technology.

Codex session ID: 019f76db-d149-7301-9131-47f898bb2d1c

Challenges

The hard part was not storing text. It was making retrieval small, deterministic, and understandable. A useful memory layer must select enough evidence to help without quietly turning the entire store into prompt payload. We also had to make deletion observable: after a record is forgotten, it cannot appear in the next recall.

What we are proud of

The product demonstrates its promise directly. Judges can add a decision, reload the page, inspect why one record was selected, ask for missing information, delete a record, and confirm it is gone. The same repository includes the tests, notebook, architecture, provider adapter, demo source, captions, and dated Build Week evidence.

What we learned

Trust improves when the system shows less, not more. A visible proof frame is easier to reason about than an invisible memory service. Refusal is also a product feature when it is tied to evidence the user can inspect.

What's next

The next step is integrating this proof-shaped memory flow directly into longer Codex sessions while preserving the same local boundary, visible scoring, hard context cap, and explicit forgetting controls.

Built With

• codex
• gpt-5.6