immutlex is an immutable knowledge base for AI agents. You ingest documents (markdown with [[wikilinks]], plus PDFs, CSVs, images, and any web page) and immutlex chunks, embeds, indexes, and links them into a navigable knowledge graph. Then it serves that graph back to your agent over MCP (74 tools) and to you over a CLI.

It runs entirely from local models in a single self-contained binary: no cloud, no API keys, no telemetry. Everything is plain markdown in S3-compatible storage you control.

This repository is the public home for immutlex: docs, releases, and issues. immutlex is a paid, closed-source product. Get it at immutlex.dev.

Why

Your agent can search the web but cannot build a persistent, cross-linked memory. It cannot remember what it read yesterday, connect ideas across documents, or find the one paragraph buried in a PDF from last week. The fix is not a bigger model, it is a substrate: the substrate is the working memory, the prompt is the question. You ask something short, and immutlex attaches the few highest-ranked chunks relevant to that question before the model ever reads the prompt. Token bills drop, recall stops being random, and confident-but-wrong answers go from common to rare.

What it does

  • Immutable knowledge base. Every document is a content-addressed revision with a full audit trail. Nothing is silently overwritten.
  • A real knowledge graph. Documents are nodes; [[wikilinks]] are edges. Forward links come from the body, backlinks are generated, and semantic edges are inferred so related docs connect even without an explicit link.
  • Fused retrieval. Semantic plus lexical search merged with reciprocal-rank fusion, plus heading search, backlink search, and graph operations (neighbors, shortest path, community detection, orphan detection).
  • Multimodal ingest. Markdown, PDFs, CSVs, images, and any web page. Frontmatter is auto-generated from the filename and body links when missing.
  • Runs local. Embedding and OCR models run inside the binary: no download step, no network call at inference time.
  • Storage you own. One S3-compatible connection string: local MinIO on a VPS, your own S3 bucket, or your Azure ADLS.
  • Cognitive maintenance. Scheduled jobs decay stale relevance, stitch orphaned docs back into the graph, and keep the knowledge base sharp over months instead of letting it bloat.

Interfaces

  • MCP server over stdio for local agents, or HTTP for remote agents. Any MCP-speaking agent (Claude, Codex, Copilot, Cursor) gets retrieval, graph traversal, and auto-injected context as native tools.
  • CLI for scriptable access from any shell or pipeline.
  • Desktop app with a 3D graph visualization, markdown editor, and live event feed.
  • Web viewer (a thin client that talks to a daemon on your own hardware, so the knowledge base never leaves your machine): https://viewer.immutlex.dev/

Add-ons

Independent tools that each emit plain markdown ready to drop in. codino and fsgdb emit it already linked (frontmatter plus back-and-forward [[wikilinks]]).

  • defuddle-rs (open source, MIT) -- point it at a URL and it strips the page chrome and returns clean markdown.
  • codino (add-on) -- pulls your Claude Code / Codex / Copilot session logs into linked markdown your agent can ask again.
  • fsgdb (add-on) -- maps a codebase (call graphs, blast radius, coverage) into immutlex-compatible markdown.

Data ownership

Open format: documents are plain markdown with YAML frontmatter in storage you control. Export the whole bucket or pull a directory of .md files and walk away anytime. No proprietary format, no vendor lock-in. What you keep by staying is everything the system builds on top of the files: the graph, the revision lineage, and the months of grounding that accrete around your corpus.

Get started

immutlex is local-first and paid. Usage-based, never per seat.

  • Pro -- $249/mo. 25,000 docs/mo and 10 GB included, single VPS you manage.
  • Pro Plus -- $499/mo. 35,000 docs/mo and 25 GB, cheaper overage.
  • Enterprise -- capacity bands, deployed Managed or Hybrid (your S3/ADLS, data never leaves your cloud), plus a fully self-hosted option.

Full details, the demo, and a hosted viewer you can point at your own daemon: immutlex.dev