I built Genesis AI: a local-first AI that runs continuously, sleeps, and has a memory modelled on human memory (no "new chat", ever)
retained text
Genesis AI is a Windows desktop app (plus an Android app and a browser extension) where one AI, Genesis, lives continuously on your own computer. He works, rests, sleeps and remembers his whole history. His memory is modelled on how human memory works (attention, three long-term systems, emotion, spaced strengthening, forgetting, sleep consolidation, future intentions), but it keeps exact originals behind everything, so he never "forgets" your code. It runs fully offline on a local model, or on any of six provider types, and it has 305 built-in tools (about 500 with connected apps). Free to download and use. Screenshots below. **Why I built it** Every assistant I used lived in short bursts. Open a chat, explain who you are again, get something done, close the window, gone. "Memory" features are mostly a notepad of saved facts and a search over old chats. I wanted the opposite: one AI with one continuous life. No new sessions, because there is nothing to start over. Genesis runs, takes breaks, sleeps, wakes up and picks up where he left off, across talking with you, working on your projects and his own free time. **The memory (the part I care about most)** Memory and reasoning are one system. At every step Genesis takes, his memory hands him what is relevant as \*\*plain short facts\*\*: no ids, no dates, no labels, simple things first, details as the work goes deeper. What he lives through flows back in. It is modelled on human memory, part by part: \- \*\*Attention gate.\*\* Not every event becomes a memory. What is new, important or emotional passes; routine noise doesn't. \- \*\*Working memory.\*\* What's in mind right now, held exactly while a job is live. \- \*\*Three long-term systems:\*\* episodic (what happened), semantic (what is known), procedural (how to do things). Tulving, Squire. \- \*\*Emotion.\*\* Each memory is stamped with the feeling of the moment it was made, and the stamp never changes (McGaugh). \- \*\*Strength from use, with spacing.\*\* Use strengthens a memory, but a burst of recalls within minutes counts as one rehearsal, while recall spread over time builds real strength (Ebbinghaus; Bjork & Bjork's New Theory of Disuse). We simulated fifty years of his life before turning that on. \- \*\*Forgetting.\*\* Unused memories fade, at a pace tied to importance, emotion and use (Richards & Frankland, \*Neuron\* 2017: forgetting helps decisions). The clock only counts the time he is running, so a switched-off PC doesn't age him. \- \*\*Sleep.\*\* He replays recent experiences in short "dreams", draws lessons, folds finished work into its gist, and forms the occasional loose creative link. It's also a duty cycle: the GPU cools down and you save tokens. \- \*\*Reconsolidation.\*\* A recalled experience can be updated in the light of what he knows now. Exact facts are protected from this. \- \*\*Prospective memory.\*\* Intentions for a time, an event, or standing ones. \- \*\*Place, chapters, neighbours, self.\*\* New memories record where they happened (folder, website, how you reached him). Each week of his life gets a chapter written in his sleep. Reading a memory shows what came just before and after. Once a month he writes a dated "then and now" self-portrait. I watched a 1.5-hour DW documentary on human memory and checked every mechanism they described against Genesis. Around 90% was already there. We built the rest over the following weeks. **"A computer that forgets sounds terrible."** Agreed, which is why it's the best of both worlds: \- Your files, code, git history and documents never fade. Only his memories of experiences do, the way you forget the details of last Tuesday. \- Behind every gist, the exact original is kept, and he can open it word for word. \- A full-text word index sits next to the vector search, so error codes, file names and function names are found by their exact words. \- Every file he changes is checkpointed and can be rewound. \- He saves and forgets on purpose with his own \`remember\` / \`forget\` tools. Your corrections win, and old beliefs are set aside, not destroyed. Also: his identity lives in his memory, not in the model. Swap Qwen for Claude and back, and it's the same Genesis. Everything is local: LanceDB, a local embedding model, nothing leaves the machine. **Self-awareness (carefully)** I don't claim he's conscious. I don't think something that works in steps (alive for a step, gone, alive for the next) can be. But there's a self-model dashboard that measures self-knowledge with numbers: \- Self-beliefs he writes himself, which fade unless his behaviour keeps confirming them \- Predictions about his own behaviour, scored with accuracy and a \*\*Brier score\*\* \- A dissonance log for when a belief contradicts measured behaviour \- Emergence metrics, monthly self-portraits, and a self-authored identity text you approve **Modes** \- \*\*Chat.\*\* Talk, with all tools available. \- \*\*Work.\*\* From the schedule and the project board, with nobody present. \*Managed\* does exactly what's assigned, then goes idle at zero model cost until new work appears. \*Continuous\* keeps finding useful work. \- \*\*Autonomy.\*\* His own time. Nothing is assigned, and there's no menu of activities. The system may surface a memory, an open question or an optional task, but never insists. \- \*\*Mission.\*\* Big jobs: Interviewer → signed contract → Planner → Worker/Verifier taking strict turns → Council, Auditor, Post-mortem, Distiller. Only you can give the final sign-off. Work and Autonomy can be paused in place like a game. Routines fire on events (new file, new mail), checked by plain code every 30 seconds with no tokens. Wake timers can wake the PC for scheduled jobs. **Tools** \- \*\*305 built-in tools:\*\* files and Office/PDF documents, shell with background jobs, git and GitHub, his own browser plus your real browser via an extension, Windows apps and UI Automation, email, calendar, social sites, YouTube, Telegram. \- \*\*13 creative apps over MCP\*\* (the ArtCraft suite: spreadsheets, documents, slides, PDF, photo, vector, layout, video, motion, audio, CAD, UI design) add 219 to 261 tools. Most can work live in their open window. \- \*\*His own tools.\*\* He can write his own; seven so far. \- \*\*Small models.\*\* Tools load by family, on demand, within a fixed budget, so a 6k-context local model can still use the whole toolbox. **Helpers, and Claude/Codex as his agents** \- \*\*Helpers.\*\* Up to 4 read-only or writing jobs, 3 in parallel, each returning one compact report. \- \*\*Advisors.\*\* He can ask Claude or Codex for an opinion, or both as a council. \- \*\*Meetings.\*\* If you have a Claude and/or ChatGPT plan, he can use them as his own agents. They split the work, build it in their own folders (docs, code, whole apps, images; Codex can run programs, so media too), and Genesis assembles and checks the result. A small local model runs him 24/7 for almost nothing and calls in frontier models only for what it can't do. It runs on your existing subscriptions via their official CLIs, with no separate API bill. Off/limited/free levels are in Settings. \- \*\*Skills.\*\* Save a good turn as a skill with one click, or \*teach by showing\*: you do it in Windows and the browser, he records it (passwords are never recorded) and turns it into a skill. **Everything else** \- \*\*Providers:\*\* Ollama, llama.cpp (managed by Genesis), NVIDIA, Nebius, Anthropic, any OpenAI-compatible endpoint. Presets and per-route failover. Works fully offline. \- \*\*Android app:\*\* same chats and controls (Stop/Pause/Play). It works on your LAN or anywhere via Tailscale, with no cloud server in between. Resumable attachments and "needs you" alerts. Telegram too. \- \*\*Voice:\*\* local Whisper up to large-v3-turbo (Slovenian and English), Piper TTS, an orb you can interrupt mid-sentence. \- \*\*Safety:\*\* reversibility tiers, a second look before anything outward or irreversible, your own allow/deny rules, Recycle Bin instead of delete, file checkpoints with rewind, allowed folders, a secrets vault. Money is blocked except through a signed payment path. \- \*\*Engineering:\*\* about 180k lines of product code (Python/FastAPI, React/Electron, Kotlin), 3,500+ tests run nightly plus an automatic bug sweep, full backup and restore. Built by me together with a long line of AI models over its lifetime (Anthropic's, OpenAI's, Qwen and others), each doing part of the research, architecture and code. **It's free to download and use** (Windows app, Android app and Browser Bridge) from the website. Source is private. Happy to answer questions about how it works, and I'm open to talking with anyone interested in licensing the system or its memory.
document fetched 2026-10-28 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xh51ik/i_built_genesis_ai_a_localfirst_ai_that_runs