LMF — Local Mind Foundation

A locally-sovereign AI assistant orchestrator for personal knowledge vaults. Designed for neurodivergent users who need cognitive prosthetics — not productivity tools.

LMF runs entirely on your hardware. No cloud dependency. No data leaves your machine. The assistant reads your vault, surfaces relevant context, and helps you think — on your terms.

---

Why This Exists

Most AI tools are built for neurotypical workflows and require trusting a vendor with your data. LMF takes a different position:

• Local only — everything runs on your machine via Ollama. No data sent to a third party.
• Vault-native — reads your Obsidian/Markdown vault directly. No import step, no sync layer.
• Conversational onboarding — the assistant learns about you through conversation, not forms.
• Operator-controlled — writes require explicit confirmation. The AI cannot modify your vault without approval.

Built for people who need the system to find them, not the other way around.

---

Architecture

┌─────────────────────────────────────────────────────┐
│                    LMF Orchestrator                  │
│  core/orchestrator.py — prompt builder, tool dispatch │
│  core/backends.py — multi-model inference (Ollama)    │
│  core/build_prompt.py — vault-aware system prompt     │
│  core/tools.config.yaml — tool definitions             │
├─────────────────────────────────────────────────────┤
│                    Features                           │
│  features/ui/ — browser chat interface                │
│  features/testing/ — test harness + analysis          │
├─────────────────────────────────────────────────────┤
│                    Deployment                         │
│  init.py — first-time setup wizard                    │
│  tools/provision-usb.py — portable USB deployment     │
│  operator/ — your local config (gitignored)           │
└─────────────────────────────────────────────────────┘

Quick Start

pip install -r requirements.txt
python init.py              # first-time setup
python core/orchestrator.py # start the assistant

Key Features

• Multi-backend inference — Ollama, any OpenAI-compatible API. Hot-swappable at runtime.
• Vault-aware prompting — reads your actual vault content (tasks, projects, memory files) into system context.
• Confirmation-gated writes — append_to_file and replace_lines require operator approval before executing.
• PII-safe testing harness — synthetic vault seeder prevents real data exposure during test runs.
• Portable deployment — USB-friendly bootstrap scripts for air-gapped or offline use.
• Test battery — 14 prompts across 4 test types: tool exercise, grounding, hallucination boundary, tool enforcement.

Test Results (qwen2.5:3b, CPU)

Metric	Score
Tool Accuracy	78%
Grounding Rate	50%
Hallucination Rate	50%
Tool Enforcement	PASS

Reference Instance

The reference deployment is Marlin — a production task surfacing engine that surfaces one task at a time via phone notification. LMF is the LLM conversation layer used by Marlin and other instances in the LMF architecture.

Related Repos

• cockpit — unified HUD for the LMF stack
• marlin — task surfacing engine
• the-time-factory — ADHD-friendly visual calendar
• prosper0 — work-specific exobrain instance

---

Part of the Local Mind Foundation architecture. Local-first, ND-designed, operator-sovereign.