Wikori Documentation

What is Wikori?

Wikori is a desktop knowledge management system that turns raw information into structured, AI-enriched knowledge. Drop any file — a PDF, a web article, an email, a voice note transcript — into Wikori and it automatically extracts the content, runs it through an AI model, and stores a richly annotated version you (and your AI agents) can search and explore.

The core idea: you own the data, it stays on your machine, and it's always in plain readable Markdown. No cloud lock-in, no proprietary format. And unlike a static document archive, Wikori is a living memory system — AI agents can write their own observations back into your vault, decisions are preserved across sessions, and search results are ranked by trustworthiness, not just keyword frequency.

Key Concepts

Vaults

A vault is an isolated knowledge base — a folder on your disk that Wikori monitors and enriches. You can have many vaults: one per client, per project, per topic. Each vault has its own ingestion pipeline, knowledge graph, and optional email tag. Switching vaults in the sidebar instantly scopes all operations to that vault.

The INGEST Folder

Every vault has an INGEST/ subfolder. Drop any supported file here and Wikori picks it up automatically — extracting text from Office documents, analyzing images, scraping URLs — then passes the content to your AI endpoint for metadata enrichment. The built-in Web Crawler can also discover and queue entire sites for bulk ingestion.

AI Enrichment

Wikori sends each piece of content to any OpenAI-compatible API endpoint. The AI extracts and returns structured YAML frontmatter: title, summary, entities, tags, source type, confidence score, and source authority. This metadata is prepended to the Markdown file and drives both the Knowledge Map and search ranking.

You choose the endpoint, and that choice is also a privacy choice. Point it at Ollama or LM Studio running on your own machine and no document text ever leaves it — not even during indexing. Point it at a hosted provider and any OpenAI-compatible service works; enrichment is a short, structured task, so the small inexpensive model tiers handle it well. There is no Wikori-branded AI service and no markup: you deal with your provider directly, or with nobody.

What Costs What

Enrichment is the only step that ever spends tokens, and it happens once per document. Everything afterwards is local and free: the index, the ranked search, the Knowledge Map and the optional vector database all live on your disk. Serving your knowledge to AI agents — any number of queries, any number of agents, every day — costs nothing. See Search & Knowledge Map for how the local layers work.

Vault Schema

Each vault automatically maintains a SCHEMA.md file — a machine-readable overview of what's inside: entity types, dominant tags, memory tiers, and a curated Ground Truth list of the most reliable entries. AI agents read this first so they understand the vault's structure before searching, leading to far smarter queries from the very first interaction.

Agent Memory

AI agents connected via MCP can write observations back into your vault — decisions, errors, patterns, preferences, and lessons. These are stored as first-class Markdown entries with full metadata, indistinguishable from human-authored documents. The vault becomes a shared, accumulating memory that persists across every session.

Search

Search reads the vault's index, so it's fast and costs nothing. Exact terms — a contract number, a tax ID — match exactly, and results are ranked by how reliable each entry is, not just how often it repeats your keyword. An optional semantic layer, running entirely on your machine, adds meaning-based matching in 50+ languages. See Search & Knowledge Map.

Automation

Each vault can run its own scheduled maintenance: process anything still waiting, rebuild the index, clear out duplicates, re-crawl a source you follow. It's the upkeep that keeps answers accurate, done without you remembering. See Automation.

MCP Server

Wikori runs a local Model Context Protocol server. Point any compatible AI agent at it and the agent can search, read, write observations, and navigate your entire knowledge base directly. Five purpose-built memory tools handle schema reading, memory saving, reliability-ranked search, session review, and memory consolidation.

Documentation Guides

Supported Formats at a Glance