Silva

Apple Notes is an excellent inbox and a poor long-term memory.

After more than a thousand notes, useful ideas were buried in an unstructured archive. Moving them into a conventional knowledge system would require extensive manual sorting and an ongoing filing habit at the point of capture.

I designed Silva to create the value of a structured knowledge base without changing how I take notes.

Synthesis above the archive

Silva leaves raw sources where they are and uses agents to maintain a compact wiki above them.

Apple Notes remains the inbox. New and edited notes are exported into an immutable raw layer, then tracked through an ingest ledger called Cairn. Agents connect related material, update existing pages and cite every claim back to its source.

The result is not another archive, but a deliberately small model of what I know.

How it works

An incremental exporter writes content-addressed Markdown files. Cairn records each item, version, ingest run, failure and retry. Agents then synthesise sources into pages about entities, concepts and questions.

Answers produced through querying are written back into the wiki, making each conversation permanently useful. Edits create new versions rather than overwriting history, and raw sources are never reorganised into the wiki.

Designed to remain useful

Silva compresses many source items into a few useful pages. Page budgets and linting prevent the wiki from growing linearly with its inputs.

Every claim references immutable ground truth through a readable locator that remains meaningful even if its original resolver disappears.

The durable system is Markdown, Git, SQLite and documented conventions. Models are replaceable labour rather than infrastructure. An AGENTS.md file contains the full operating contract, including provenance rules, page types, citation formats and limits on autonomous changes.

Any capable future agent should be able to tend the system by reading one file.

Private by construction

Silva is local-only and has no public endpoint. It is designed for read-only access over a private Tailnet.

Cairn uses SQLite operationally and can emit human-readable JSONL snapshots for recovery, while the wiki retains full Git history.