informational · 2026 · active

Verdant

Problem

Conversations with language models are high-density thinking sessions that evaporate the moment the tab closes. Cloud chat clients discard structure; they store transcripts, not knowledge. And for anyone working with sensitive material, uploading those transcripts is not an option at all.

Verdant asks: what if the conversation itself were raw material for a knowledge base — extracted, structured, and queryable, entirely on your own machine?

Architecture

Verdant is a desktop application built on Tauri v2 with a Rust core and a React/TypeScript frontend.

Rust backend (Tokio, Rusqlite, Serde): multi-threaded command handlers for message retrieval, graph extraction, and system hooks. Everything persists to a single local SQLite database.
Provider layer: connects to Ollama and custom local endpoints, auto-detecting connection status and active models. No cloud path exists.
Extraction pipeline: as sessions accumulate, Verdant parses them, detects key entities and concepts, and synthesizes them into a graph.
Frontend: React with an interactive knowledge-graph view (React Flow), a persistent Memory Bank with categorized memories, and a full chat interface with markdown, syntax highlighting, and LaTeX rendering.

Decisions

Local-first as a hard constraint, not a feature flag. Choosing Tauri + SQLite over Electron + a synced store made offline the default state rather than a degraded mode.
Structured memory over raw transcripts. Memories are typed (facts, instructions, project scopes) and independently editable — the extraction is a starting point, not an oracle.
Graph as a first-class view. The knowledge graph is not a visualization bolted onto search; it is the primary navigation surface for accumulated context, with live extraction available inside the chat window.
Export as escape hatch. Sessions, memories, and graph state export to Markdown/JSON — the database is never a lock-in point.

Results

A working, self-contained desktop workspace: multi-session chat against local models, autonomous knowledge-graph construction, and a memory bank that persists context across sessions — with zero bytes leaving the machine.

This is the "knowledge infrastructure" thread in its purest form: raw personal data (conversations) turned into navigable structure.