Your AI assistant forgets everything about you between sessions. Daniel Miessler thinks that's the core problem holding back the entire category.
His open-source project, Personal AI Infrastructure (PAI), has quietly grown to over 10,000 GitHub stars and reached version 4.0.3. The central argument: chatbots like ChatGPT and Claude operate in a stateless loop of query-answer-forget, and even agentic tools like Claude Code lack real personal context about the person using them. PAI tries to fix that with persistent identity, memory, and a continuous learning cycle.
The TELOS Identity System
The most distinctive piece is TELOS, a set of ten structured files that capture who you are: your mission, goals, beliefs, mental models, active projects, challenges, and ideas. The premise is simple but rarely attempted - an AI that knows your five-year goals, your working style, and the things you've already tried and abandoned will give fundamentally different advice than one starting from zero every conversation.
This isn't a vague concept. The TELOS files (MISSION.md, GOALS.md, PROJECTS.md, BELIEFS.md, and six others) are plain text documents you maintain yourself. The AI reads them before every interaction.
Architecture Over Model Choice
PAI follows 16 principles, and the most useful one for practitioners is "Scaffolding > Model" - the system architecture around your AI matters more than which model you pick. Another worth internalizing: "Goal, then Code, then CLI, then Prompts, then Agents," a strict hierarchy that says you should reach for a bash script before a prompt, and reach for a prompt before an agent.
The memory system uses three tiers (hot, warm, and cold storage) and logs every interaction for signals like ratings, sentiment, and success/failure patterns. There's a hook system with eight lifecycle events, a security layer that validates commands before execution, and voice integration through ElevenLabs.
Who This Is Actually For
Miessler explicitly targets "everyone universally" - small business owners, artists, non-technical people - not just developers. The open-source license is intentional; he frames proprietary AI assistants as gatekeeping.
The gap between PAI's ambition and current reality is still wide. Setting up ten identity documents, maintaining a three-tier memory system, and configuring lifecycle hooks requires technical comfort that most of the stated target audience doesn't have. The AI-powered installer helps, but the maintenance burden is real.
Still, the core insight is worth stealing even if you never install PAI: write down your goals, projects, and decision-making principles in a format your AI tools can read. A simple GOALS.md file that you paste into Claude or ChatGPT at the start of important conversations will get you 80% of the benefit with 5% of the setup. The structured identity approach is the genuinely good idea here, and it works regardless of what tools you use.