Qorx Zero

Inspiration

Every AI tool has the same quiet failure: memory. You either re-paste the same project context into every session, or you hand your entire history to a remote memory service and hope it sends the model something relevant. Both are bad. The first wastes tokens and time. The second ships your whole memory store off-device and gives you no say in what the model actually sees.

We wanted memory that stays with the work — on your device — and that sends the model only the proof it needs for the task in front of it.

What it does

Qorx Zero is a persistent, cross-session MemoryAgent:

Device-local store. Every decision, preference, and note lives in IndexedDB in the browser. Nothing leaves the machine except the bounded proof frame for the current question.
Transparent retrieval. For each question an inspectable ranker scores records — 0.65 × keyword overlap + 0.20 × importance + 0.15 × recency — and selects only what's relevant.
Timely forgetting. Records carry a TTL and can be forgotten explicitly; expired memory is dropped before it can reach the model.
Bounded context. The proof frame is hard-capped at 5 records / 1,600 characters, so recall stays inside a limited window no matter how large the store grows.
Evidence-gated answers. Qwen answers only from the proof frame and cites source hashes. Ask something unsupported — e.g. a password we never stored — and the frame is empty and the agent refuses to invent an answer.
Learns locally. Successful, proof-supported Qwen interactions become deduplicated 30-day local experiences, so the agent improves across sessions without uploading its memory.

How we built it

Frontend + store: Next.js / React, with the full memory store in IndexedDB.
Ranker & proof frame: a small, dependency-free scoring function so every recall decision is inspectable, not a black box.
Model: a thin server adapter (app/api/ask/route.ts) calls Qwen via Alibaba Cloud Model Studio (DashScope). Only the capped proof frame + the question ever reach the API — never the full store.
Deployment: packaged for Alibaba Cloud Function Compute as a custom container (deploy/alibaba/).

What we learned

Bounding context beats hoarding it. The hard cap on the proof frame forced better retrieval — a small, well-chosen frame outperformed dumping everything.
A refusal is a feature. Making the agent say "no proof, no answer" instead of hallucinating turned out to be the most trust-building behavior in the whole demo.
Keeping memory on-device changes the privacy story entirely — the model can be useful without ever holding your history.

Challenges we ran into

Retrieval quality vs. context limits: tuning the ranker weights so the right 5 records surface within the character cap.
Honest forgetting: making sure TTL-expired and explicitly-forgotten records are excluded from recall before the frame is built, not after.
Clean-room discipline: keeping this edition free of any private engine — it's a complete, standalone app with no hidden dependencies.