The week Agent Memory got its name

Twenty-one heuristics in one day

On 19 April, twenty-one numbered Heuristics pages went down in a single day. Build from current practice. Wiki lint or drift. Privacy at the write. Operator doesn't own the data plane. Each one a lesson earned, captured so future work wouldn't repeat the past wrong turn.

The cascade wasn't bloat — it was a corpus that had been accumulating informally since Day 1, finally getting filed. Each heuristic carries a one-line hook, an origin (how it was earned), and a how-to-apply note. The set has grown since (currently 36 numbered) and pruned (two retired in May). The discipline of filing rather than accumulating is what makes the set load-bearing.

A heuristic that lives only in the operator's head is a heuristic that doesn't fire.

Agent Memory

Concept 29 — Agent Memory — landed on 18 April. The brand metaphor for what the Neuro Agent delivers: a shared holding space between the human and the Agent. The human captures; the Agent holds, organises, and surfaces.

The naming was the thing. Externalised executive function system is clinically accurate but doesn't sing. Agent Memory is two words a prospect can carry. It's also unowned in the assistive-tech and ND-support space — distinctive in the way prosthetic memory (Concept 23) is distinctive. Months later, Agent Memory is on the homepage; externalised executive function system still carries the funder register.

Two words a client can carry beat a clinical phrase a clinician can defend.

Library, Support, Home

Concept 27 named the workspace architecture: a Library with two spaces. Support (later renamed Neuro Support) holds the executive function scaffolding Jamie can see. Home holds the diary, project knowledge bases, anything the client wants their Agent to remember but no one else to see. The client's Claude routes captures by interpretation, not by tags.

The architecture made interpretation-based privacy operational — Heuristic 03 in its working form. Private content never lands in operator-visible storage because the Agent extracts only the support-relevant slice on the way in. Privacy enforced at the write, not the view. There's nothing to leak because there's nothing there.

Privacy you can configure is privacy that can leak. Privacy you can't write isn't there to leak.

Soul and skeleton

Implementation 60 named the architectural split: identity (name, personality, behavioural rules) lives in a dedicated file separate from operational infrastructure. Soul and skeleton. CLAUDE.md is the skeleton; soul.md is the soul. They auto-import together at session start.

The split mattered because identity drift is invisible until it's expensive. Without a named soul file, the agent's voice migrates a fraction per model update; six versions later, the project sounds like someone else's project. With a named soul file, the drift has a target — edit the soul, not the skeleton. Heuristic 21 (model updates shift voice) made the failure mode explicit a day later.

Identity is infrastructure. So is skeleton. Mixing them costs the next migration.

What's next

Week 5 brought the architecture into running form — the first Cloudflare Worker shipped (calendar sweep), the Home page took its working layout, Privacy as Absence was named, the block-track architecture set the pricing structure, and the Notion Capture Shortcut shipped.

The discipline had a corpus. The architecture was about to start running.