ECHOESWITH EVIDENCE
Your stories. Your device.

A LIVING ARCHIVE / A QUIET OFRENDA

Some things
deserve to echo.

Keep the small stories. Find the threads between them.
Every connection leads back to the words you saved.

Memory, with a source. Never a made-up answer.

FOLLOW A THREAD

What are you looking for?

Not a chatbot. A way back to your sources.
TRY A THREAD
An open-source mind. A private archive.

Load a small AI model to search by meaning, right here on your device. First download: ~23 MB of model weights plus browser runtime files from Hugging Face and jsDelivr. Your stories are never sent.

AI not loaded. Search currently uses keyword matching (NOT AI).

No AI download until you choose.

THE LITTLE THINGS, KEPT

Your memory table 08

The eight starter stories are entirely fictional — a small, imagined community. Add your own, or explore theirs.

A home on this device.

By default, edits last for this tab. Export a copy before you leave.

ADD TO THE TABLE

Keep a small story.

Saved as text on your device. Nothing is uploaded. Search works best with short English memories. Only the first 256 wordpieces of a card or query are embedded; the full story remains available to read. The model can rank the wrong story.

THE GHOST IS A CONNECTION, NOT A VOICE

We find. We don't pretend.

Echoes is a small virtual ofrenda: a place to honor what came before, without pretending that an AI is a person who has passed away.

  1. Keep a story. Add a memory in your own words. The demo stories are fictional.
  2. Choose local AI. Only the load button downloads the open-source model and runtime. The providers see ordinary download metadata such as your IP address. They do not receive your story or query text.
  3. Follow the evidence. The model turns your query and cards into 384-number vectors, compares their cosine similarity, and returns exact source excerpts. Similarity is a ranking signal, not a probability or a fact check.

The model is downloaded from Hugging Face and the runtime from jsDelivr. Your browser may cache those files; future loads may still require network. This page has no analytics, accounts, paid API, or server database. Story cards remain in tab memory unless you opt into this browser's local storage or export a file. Other people using the same browser profile could read locally saved cards.

If AI cannot load, you can keep using the clearly marked keyword search. No generated answers, no impersonation, no hidden AI simulation.

OPEN TO INSPECTION

Small model. Real work.

Embeddings: Xenova/all-MiniLM-L6-v2, Apache 2.0, ONNX q8. Based on sentence-transformers/all-MiniLM-L6-v2.

Inference: Transformers.js 3.8.1, Apache 2.0, using ONNX Runtime Web (MIT). Runs in a Web Worker on this device.

Original interface, CSS illustrations, application code, and fictional sample stories created for the October 2026 challenge with AI assistance. No external images, fonts, or tracking.

Third-party notices · Project & demo notes