# WisprLink

> WisprLink is a real-time memory layer for writing. As you type or dictate, it links phrases to the exact files, images, URLs, people and glossary terms they refer to, from a personal "Memory Palace". Your original text is never changed; a SHA-256 hash proves it.

WisprLink is a concept workbench (MVP), not a production service. It is English-only today; multilingual support is planned for launch.

## What it does
- Linking: underlines phrases such as "hero image" or "Priya" and links them to the matching memory, with a hover card explaining what it is, why it matched and how confident the match is.
- Vague references: resolves phrases like "that photo" or "the file I selected" from a working context.
- Learning: clear statements like "The staging bucket is s3://acme-stage." are saved to the Memory Palace in the same request, so earlier text can be linked too (backfilling).
- No-match tracking: checks that find nothing are recorded and revisited as more detail arrives.
- Context Pack: exports the text with every link expanded, ready to paste into an AI assistant or IDE.

## How it works
- The browser decides when to check: at least 25 words in total, then after a one-second typing pause, or straight away for large dictated or pasted chunks. Text that is vague (low specificity) or drops in specificity triggers a check.
- The server matches with exact, fuzzy, phonetic, BM25 and local semantic (n-gram TF-IDF) signals, fuses them with reciprocal rank fusion, and surfaces only confident links per profile.
- Optional AI reranking only for ambiguous links when a key is configured; otherwise local matching works end to end.

## Who it is for
- Builders who re-type file paths and project facts into every AI chat.
- Dictation-first writers (for example Wispr Flow users).
- Product and research leads writing briefs that reference docs, people and assets.

## Future work
- Front-runner models: a light quantized or Siamese model gives instant previews while typing; a heavier model confirms on pause.
- Multilingual lexicons, fact patterns and cross-lingual embeddings.
- Incremental processing (query correlation) so only new text is re-scored.
- Learned confidence calibration from feedback, benchmarks and an A/B arena.

## Links
- App: https://wisprlink.amaagra.com/
- Docs: open the "Docs" tab in the app for the full product and engineering write-up.