Boswell wasn't designed from a whiteboard. It was designed from nature.

// THE ARCHITECTURE

One governed memory your whole fleet runs against.

Every client — Claude, ChatGPT, Gemini, your own local models — connects over MCP to a single tenant-scoped memory. Writes don't get stored on faith: a small fleet of distilled models curates, fact-checks, and gates each memory before it's kept. Reads come back as recall, not a flat list.

Tenant-isolated (row-level security) · model-agnostic over MCP · curation on owned hardware (no per-token cost) · measured 95/100 vs the prior cloud-curated corpus on a blind faithfulness A/B.

// THEORY 01 — EMERGENCE

Intelligence lives in the trails, not the nodes.

A brainless slime mold — Physarum polycephalum — placed on a map of Tokyo with food at each rail station, grew a nutrient network that matched the engineered rail system. 500 million years of evolution solving the same optimization our engineers did.

That's the most important idea in Boswell: meaning isn't in the memories, it's in the paths between them. Boswell stores trails as first-class objects. They thicken toward what you use and thin toward what you don't — so the graph grows a shape that looks like your actual thinking.

// THEORY 02 — DECAY

Forgetting is health, not failure.

A mind that weighted everything equally would drown. Human memory runs on a forgetting curve, and Boswell decays on the real one — FSRS-6, the spaced-repetition model — so what matters stays bright and the noise goes quiet.

But not everything decays equally. Facts fade; decisions and the reasons behind them persist. What was at stake is what survives — because that's what you'll need to reach back for.

// THEORY 03 — RECOGNITION

Recall should feel like recognition.

Around 300 milliseconds after a mind registers something meaningful, the brain fires a measurable wave — the P300, an event-related potential. It's the electrical signature of “oh — this matters.”

Boswell's recall is engineered to fire it in sequence: the answer first, then the adjacent memories you didn't ask for. The delight is in the unpacking, not the answer. A vector database returns text. Boswell recreates the experience of remembering.

This isn't borrowed language. Boswell's founder, Steve Krontz, is a published brain–computer-interface researcher — Randolph, Warren, Krontz & Pate, “Using Neural Input to Control Google Glass,” NeuroIS 2014 — and used the P300 evoked response to give a patient who couldn't speak a way to talk through a screen. Director, KSU BrainLab.

// THEORY 04 — CONSOLIDATION

The brain bookmarks awake and consolidates in sleep. So does Boswell.

The hippocampus encodes candidate memories during the day and replays the important ones during sleep — sharp-wave ripples, the mechanism the Buzsáki Lab mapped in Science. Boswell implements both halves: a fast encode at commit (the P300 moment), and a slower consolidation pass that promotes what proved to matter and lets the rest settle.

Memory isn't a single write. It's a two-phase circuit — and Boswell is one of the few systems that models both ends of it.

// THEORY 05 — CONSCIENCE

Ethics isn't a filter bolted on top. It's the shape.

Boswell is built on an original model of cognition — a cone whose layers converge from instinct → emotion → reason → conscience → ethics. The claim is structural: a mind shaped this way can't produce unethical output, because the geometry won't allow it to reach the tip without passing through conscience first.

In the system, that's the immune layer: suspicious content is quarantined, never deleted — the immune system proposes, the mind disposes. Boswell watches for you, never over you. Memory with a conscience, built by someone with one.