ML Architecture

Ambient Intelligent Systems

August 16, 2026agentsembeddingshnswswiftgrpcp2pknowledge graphs

How three programs taught each other a paradigm.

Bonnie acts, Seer reasons, Totem remembers connections.

That sentence is the shortest true description of the stack I've spent the last ten months building — a reasoning server, a memory node, and a macOS assistant, three programs composed into one system. This book is the story of how they got that way, told through the actual decisions and the actual notes, ending at the thing the journey forced me to name: Ambient Intelligent Systems — an architecture paradigm for AI that works with the user, not for the user. Parallel to the human. Supporting, not replacing. Built on infrastructure the person owns.

The industry's default answer to "personal AI" is a chat window with a memory file — and, as one of my notes puts it, systems like that remember what you said; they do not learn how you think. The stack described here is the other answer, and the paradigm is what its parts turned out to have in common.

Timeline from Seer's start to the AIS paradigm — Figure

How to read this

Nine chapters in three parts, written to be read in order — each early chapter earns exactly one principle that the final chapter assembles — but every chapter stands alone.

Part I — The Server is Seer: the personalized-retrieval loop, the partition insight that moved learning to the granularity of ideas, and the mesh whose one unsolved problem moved the whole architecture onto the user's machine. Part II — The Node and the Join gives memory a body (Totem) and forks every turn into a lane that speaks and a lane that acts. Part III — The Paradigm is the destination: the ambient layer of worlds and realms, capability as bounded data, the external work that independently arrived at the same laws — and the name. Two appendices collect the field notes and the lexicon.

Scattered through the text are field notes: dated blockquotes taken from my real working notes, lightly trimmed and never rewritten. They are artifacts of their moment — several predate the paradigm having a name, and some record numbers that were true only on the day they were written; the surrounding text says so when it matters. Codenames (Seer, Totem, Sinatra, Oracle, Gita, Conduit) are the real ones.

written August 2026 · the story runs 2025-10-23 → present

Contents

01Seer: the personalized-retrievalThe project begins (2025-10-23) with a measurable claim: personalization lives in retrieval, not generation. Sinatra, Oracle, Gita — and the candid review that became the roadmap.02Partitions: paying for ideas, not containersThe April insight: a partition is a thought, content-addressed across every document it appears in — and the loop that closes only at that granularity.03The mesh, and the cost of scaleA two-level HNSW over gossiped tag embeddings, the rare-topic recall problem — and the demotion that moved the assistant's center of gravity onto the user's machine.04Totem: memory becomes a nodeMemory as files the user owns, a UUID that is the identity, and the three-process topology joined by one wire contract.05Bonnie: one turn, two lanesEvery turn forks: a lane that speaks and a lane that acts, structurally forbidden from sharing a prompt. The freeze, the guardrail, the receipts.06The Ambient architectureThe destination: worlds and realms, the calendar read that reached nobody, and the doctrine that eyes are an upgrade — never a precondition.07Abilities: bounded data, frozen turnsCapability as validated, portable data instead of persuasive prose — and a mechanism-by-mechanism comparison with Anthropic's SKILL.md.08Composability, proven elsewhereA 2026 preprint on spatiotemporal composability and DeepSeek's harness arrive at the same laws independently — and the exact place an ambient system must go further.09Ambient Intelligent SystemsThe paradigm named: seven properties, each earned by a chapter — and where this book deliberately stops.10Appendix: the field notesEvery note, in chronological order, with links to where each is woven in.11Appendix: the lexiconThe load-bearing nouns, spelled exactly.