Fortemi: A Rust-built Private AI Knowledge Base That Truly 'Understands' Content

Fortemi is a self-hosted AI knowledge base built with Rust and PostgreSQL, aiming to bridge the gap between data storage and content understanding. Its core features include hybrid semantic search, automatic knowledge graph construction, multi-modal content processing, MCP protocol support, and privacy-first design. It enables deep understanding and intelligent association of various content types (documents, images, audio, video) while ensuring full data control via privatization.

Traditional knowledge management systems act as mere storage warehouses—they save data but fail to understand content meaning. Keyword-based search requires users to remember exact terms (e.g., searching 'how to use AI to answer document questions' won't find notes using 'retrieval-augmented generation' or RAG). This misalignment with human conceptual memory (remembering 'what' but not 'exact words') creates a significant gap Fortemi aims to solve.

Fortemi (pronounced for-TAY-mee) follows the core concept 'Memory that understands'. It's a self-hosted system written in Rust (~160k lines of code) with PostgreSQL backend and Node.js-based MCP server. It runs on consumer GPUs (8GB VRAM) without cloud dependency, allowing full privatization. It not only stores content but also understands its meaning, discovers concept relationships, and builds contextual connections.

Hybrid Semantic Search: Combines BM25 (exact text match), dense vector similarity (semantic understanding), and RRF fusion (intelligent result merging) to find relevant content even with different terminology (e.g., 'AI answer doc questions' → RAG notes). Supports multi-language (CJK included) and emoji understanding. Auto Knowledge Graph: Grows organically with content: auto-links (similarity >70%), SNN scoring (relevant pairs), PFNET (sparse redundant links), Louvain (concept groups), SKOS (hierarchical organization). No manual tagging needed.

Multi-modal: 13 adapters handle visual (image OCR/description, video keyframes/transcription, 3D model rendering), audio (Whisper transcription, pyannote speaker separation), docs/mails (email parsing, spreadsheet analysis, zip decompression). Derived attachments (thumbnails, subtitles) make media searchable. Content-aware: Detects 131 doc types with tailored strategies: code (grammar-aware chunking), prose (semantic chunking), meeting notes (extract decisions/actions), research papers (focus on methodology/findings).

MCP Protocol: Implements Model Context Protocol server with 43 tools for AI agents (search knowledge base, create notes with links, query knowledge graph, process multi-media, maintain dialogue history). Acts as a central knowledge hub for AI workflows. LLM Support: Unbound to specific models; supports Ollama (default), OpenAI, OpenRouter, llama.cpp. Hot-switchable configs and GPU concurrency control for multi-user scenarios.

Deployment: Two ways—1. HotM desktop app (Linux/macOS/Windows, no Docker/PostgreSQL config, for individuals); 2. Docker self-hosted (for teams/developers, supports air-gapped environments with profile options based on VRAM). Privacy: Full privatization (local data/models), X25519/AES-256-GCM encrypted sharing, OAuth2/API key auth, multi-memory archives (schema-isolated spaces for multi-tenancy). Sensitive data never leaves local servers.

Fortemi evolves knowledge management from passive storage to active understanding. It's suitable for personal note management, team knowledge bases, and AI agent context support. As an open-source project, it represents a key direction in intelligent knowledge systems—where value lies in connecting and discovering knowledge rather than just storing it.