◦ Prompt · Mind

Give Your AI Agent a Brain and a Memory — From a Stateless Chatbot to an Assistant That Stays Herself and Remembers You, in One Session

Paste this into Claude Code, Cursor, or Aider and it interviews you about the agent you already have — its stack, whether it remembers anything today, whether its voice drifts over long chats — then does a strictly read-only pass over your codebase to see how it actually thinks before it changes a line. Only then does it help you build the brain that runs Trillion: a personality that lives in an editable file and reloads the moment you save it, a two-block system prompt that keeps the whole personality resident every single turn without ballooning your bill, an always-loaded core-knowledge layer so it never re-asks what it should already know, working memory that persists and recovers across sessions, and a typed, file-backed long-term memory with semantic recall — written both by the agent's own hand through tools and by an automatic end-of-session extractor that dedupes before it saves. It finishes with a self-knowledge layer so the agent knows what it is, and a personality checkpoint that stops it sliding into generic-assistant voice deep in a conversation. Interview first, map second, then build tier by tier — each tier ships independently with its own verification.

Jul 3, 2026agentbrainmemoryrecallpersonalitycontext-engineeringclaude-apianthropic-sdktutorial

◦ 4,386 builders waiting

Want the agent behind these prompts?

These prompts came out of building Trillion. The whole thing is going open-source on GitHub. Drop your email and I'll ping you the moment the repo drops.

◦ Next prompt

Give Your AI Agent a Board of Advisors — Turn Well-Documented Operators Into Seats That Actually Disagree With You, Cite Their Sources, and Refuse to Answer When They Don't Know, in One Session →

Paste this into Claude Code, Cursor, or Aider and it interviews you about the agent you already run — what it can read about your business, how it surfaces things to you, whether it can schedule its own work — then researches each advisor you name, fact-checks that research against itself, and builds the machinery: one markdown dossier per seat holding numbered, citable doctrine; a router that picks who has standing on a given question; a fan-out that gives each seat its own isolated model call so they can genuinely disagree; a citation gate that strips any reference a seat could not have seen; and your own agent as chair, the only participant who has read your actual revenue. Plus abstention as a first-class answer, a unanimity guard that cannot be talked into agreement, blind spots that let the chair discount a seat, and a standing monthly review that puts an agenda in front of you without being asked. Interview first, research second, then build tier by tier — each tier produces something you can run, and the anti-fabrication defenses are built before the first meeting, not after.