The archive · Product Ideas · Product decision · 2026
LocalGPT: a Rust AI assistant whose plain-file memory also dreams explorable 3D worlds
A single-binary, local-first AI assistant keeps memory in readable markdown files and learns your style — then generates explorable 3D worlds saved as skills.
LocalGPT
What it had to solve
AI assistants usually keep their state in opaque cloud accounts and their chat output disappears when the session ends. The LocalGPT project set out to build a local-first assistant whose long-term memory, task queue and personality are plain files on the owner's machine, and paired it with a second binary that turns the same engine toward generative 3D world-building.
How it works
LocalGPT arrived on Hacker News as a Show HN on February 8, 2026, billed as 'a local-first AI assistant in Rust with persistent memory.' The repository's tagline is two ideas at once: a local AI assistant, and an engine that dreams explorable worlds — and the project's own blog entry, linked from the README, says the whole thing was built in four nights.
The assistant is a single Rust binary with no Node.js, Docker or Python dependency. Its state is deliberately human-readable: a workspace holding MEMORY.md for long-term knowledge, HEARTBEAT.md as an autonomous task queue, SOUL.md for personality and behavioral guidance, and a knowledge folder — all indexed with SQLite FTS5 for keyword search and sqlite-vec for semantic search using local embeddings. It runs as CLI, daemon with an HTTP API and web UI, desktop GUI or Telegram bot, and can point at local models through LM Studio or Ollama or at cloud providers.
The second binary, localgpt-gen, is an AI-driven 3D world-builder on the Bevy game engine. Prompts create scenes with parametric shapes, PBR materials, point, spot and directional lights, behaviors such as orbit and path-follow, and ambient or spatial audio, exported as glTF/GLB, screenshots or browser-viewable HTML. A headless mode queues experiments for overnight batch runs, and the memory system learns the user's creative style — palettes, lighting preferences, entity templates — and applies it to future generations, while complete worlds save and load as reusable 'world skills.'
The February 2026 Show HN drew 331 points and 153 comments; by the September 2026 capture the repository had passed 1.1k stars with 101 forks and 619 commits. Critics on the thread objected that the name overstates locality while cloud LLM providers are supported, which the README answers by documenting fully offline use through LM Studio and Ollama alongside optional cloud providers. The deeper bet is persistence applied to output: if worlds can be saved, reloaded and learned from, the assistant's creations accumulate the way its conversations do.
Why it lands
- Plain-file memory — MEMORY.md, HEARTBEAT.md, SOUL.md, knowledge — makes the assistant's state inspectable and editable, answering the opacity of cloud assistants directly.
- One Rust binary with no Node, Docker or Python makes local-first genuinely low-friction to install and run on the owner's own machine.
- The style memory learns palettes, lighting and entity templates across sessions, turning world generation into an evolving personal medium rather than one-shot prompts.
- Saving complete worlds as reusable skills closes the loop: the same persistence that remembers conversations also preserves creative output, so nothing the machine makes is ephemeral.
What it did
The Show HN post drew 331 points and 153 comments on February 8, 2026, with discussion centering on what 'local' means when external LLM providers are supported. By the September 2026 capture the Apache-2.0 repository showed 1.1k stars, 101 forks and 619 commits, and the project's distinctive move — worlds as skills the assistant remembers and reuses — made its creative output persist exactly like its memory.
What you can take
Make the machine's state legible: memory as plain files turns a personal AI into something the owner can audit, and pairing it with playful 3D world generation turns a utility into a creative tool.
Since then
As of the September 2026 capture, LocalGPT remains open source under Apache-2.0 with 1.1k stars, 101 forks and 619 commits, and the Show HN post held 331 points and 153 comments. Hacker News commenters argued 'local' overstated the case while external providers are supported, so the README documents fully offline use through LM Studio and Ollama alongside cloud options. The maintainers keep shipping the assistant and world-builder as separate crates, with a headless experiment queue, an MCP server for desktop AI tools, and posts framing explorable worlds as agent skills.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe the brief you are staring at, and see who has been given the same problem.
Free account · 3 free questions · no card