wiki0

Local-first AI memory as a Markdown wiki backed by SQLite search and
MCP tools.

Value prop

wiki0 gives your AI an explicit, inspectable memory: a Markdown wiki
it can read and write, backed by a disposable SQLite index for fast
search, backlinks, facts, and graph views. No opaque vendor memory;
the knowledge stays local, diffable, and portable.

It is meant to be memory/context infrastructure rather than another
opaque chat history: agents can preserve decisions, retrieve cited
project context, inspect sources, and update the underlying Markdown
so future sessions start from durable knowledge instead of vibes.

Shape

wiki0/
  wiki/             # canonical Markdown knowledge
  .wiki0/           # local SQLite index/cache
  packages/core/    # schema, indexing, wikilinks, search primitives
  packages/cli/     # wiki0 init/index/search/context/lint/graph/facts
  packages/mcp/     # MCP server wrapping core operations
  apps/web/         # SvelteKit/mdsvex human interface

Planned workflow

wiki0 init ~/my-wiki
wiki0 plan --source_type codebase --scope "current repository"
wiki0 extract docs/guide.md
wiki0 sync docs --root ~/my-wiki
wiki0 index
wiki0 search "why sqlite"
wiki0 context "what did we decide about memory?"
wiki0 lint
wiki0 graph
wiki0 topics
wiki0 events
wiki0 facts add "SQLite is a rebuildable index" --category decision

Agents use the MCP server to create pages, link related knowledge,
retrieve context with citations, and queue uncertain claims for
review.

Design stance

• Markdown is source of truth.
• SQLite is a rebuildable index/cache.
• [[WikiLinks]] follow an Obsidian-compatible subset.
• Git/cloud sync moves the wiki between machines; the index can be
  rebuilt.
• Web UI is optional, for browsing, review, and graph visualisation.