All thoughts

Agent OS, Part 1: Foundations

Stages 1 and 2: why I stopped chasing model launches, then giving the model a memory.

How it works

Day 1 · Dec 2025 Stop chasing launches

Day 1: one agent, one workflowA single platform with one block on it: one thin agent for one workflow I own. The block is lit in amber.One agentone workflow

Foundation-model competition got boring. What matters is whether something saves time, makes money, or ships under real constraints.

Attention is the scarce resource (yes, the Transformer pun). Agency beats credentials. Vibe coding gets you about 90% of the way, and the last 10% of shipping is the hard part.

Day ~175 · Jun 2026 Give it a memory

Day ~175: the agent gets a memoryA memory platform now sits under the agent. On it, a stack of three sheets for the procedural, semantic and episodic pillars, a pile of loose sheets for the raw chat log, and a small consolidation gate between them, lit in amber.Memory3 pillars + gateAgentone workflow

An LLM is a powerful brain that doesn't know you. Memory is how you make it useful. I split it into working memory plus three pillars:

  • Working: context RAM for this turn
  • Procedural: how to act (skills, SOUL.md, playbooks)
  • Semantic: durable facts (profile, RAG)
  • Episodic: dated events (chat log, state.db)

Don't dump the whole log into every turn. Put a consolidation gate between episodic and semantic: after N chats, a cheaper summarizer distills the facts. Token efficiency is product design.

My take

What I learnedApps beat model theater.

What I learnedMemory is the moat.

Try this

  1. Pick one workflow you own. Ship a thin app or agent that changes it, and ignore the next model drop until it ships.
  2. Label every fact you store as procedural, semantic or episodic. Refuse "memory" that's just a growing paste.

Sources