The Promise
Build this once and your AI stops re-meeting you every morning — it walks in already knowing the business, the team, the open decisions, and the constraints that haven't changed since last Tuesday.
The One-Sentence Setup
Most operators think their AI gets smarter the longer they use it — it doesn't, because LLMs have no memory; what compounds is the file system the operator engineers around the model.
The Core Insight
An LLM has a context window, not a memory. Every new chat is a goldfish that just got handed a notepad. Memory is something the operator builds on top of the model: a structured set of files the AI reads before it answers, and writes to during the session. The compound effect doesn't come from the model getting smarter — it comes from the operator's notes getting longer, sharper, and more referenced. Four file types, one folder, loaded into every session. That is the entire architecture. The reason most stacks feel dumb on day 90 is the operator never built the folder.
The Mechanism
1. The Identity File — who the operator is
What: A single file capturing role, business, family, fixed constraints, non-negotiables. Things that change once a quarter at most. How: One page of markdown. Plain prose, headings for sections. Examples: "I run a 14-unit franchise group." "I have two kids under ten — calendar holds 5-7pm school nights." "I do not do equity-for-services deals." This file is the AI's character sheet for the operator. Miss this and: the AI keeps suggesting Tuesday 6pm calls and recommending strategies that contradict the operator's actual position.
2. The Context File — what's happening right now
What: Active projects, this quarter's goals, this week's constraints, decisions currently in motion. Changes weekly. How: Treat it like a standing agenda. Top of the file: "What I'm working on this week." Middle: open loops and decisions pending. Bottom: the three numbers that matter right now. Rewrite Friday afternoon — five minutes, not thirty. Miss this and: the AI is technically informed but tactically useless, recommending things that contradict what the operator is actually doing this month.
3. The Decision Log — append-only history
What: Dated entries of meaningful calls the operator has made. The decision, the alternatives considered, the reasoning, the expected outcome. How: Newest at top. Date-stamped. Never delete, only append. Format: Date · Decision · Alternatives · Why · Revisit-by. The AI reads this to know what's already settled. When the operator asks a question the log already answers, the AI says "we decided this in March — here's why" instead of re-litigating. Miss this and: the AI re-opens closed decisions every week, and the operator wastes the highest-leverage hour of the day re-justifying calls they already made.
4. The People Profiles — one file per person
What: A short profile for each meaningful counterparty — direct reports, key partners, investors, board, top customers.
How: Title, role, last interaction, communication preferences, sensitive topics, current open thread. One file per person, kept in a People/ subfolder. The AI references the relevant profile when drafting outreach, prepping for a call, or interpreting a message.
Miss this and: the AI drafts emails that read generic, and the operator either rewrites every one or sends something that sounds like a vendor template.
5. The Load Pattern — get the files into the session
What: The mechanism that puts Identity + Context in front of the model on every chat.
How: Three options, pick one. (a) Claude Projects: drop the files in the project's knowledge sidebar. (b) Claude Code: drop a CLAUDE.md at the repo root and the agent reads it automatically. (c) Cursor: put the same files in .cursor/rules. Turn on Anthropic prompt caching with cache_control: ephemeral so the same 4,000 tokens of memory cost 10% of the first-call price on every subsequent message.
Miss this and: the architecture exists but never gets read, which makes it identical to not having one.
6. The Append Loop — keep the files alive
What: The discipline that adds new entries during or after every meaningful session. How: Two paths. (a) Manual: end-of-day voice memo → Whisper transcript → paste into Decisions.md. (b) Automatic: Claude Code writes the entry at session close with a single prompt — "append today's decisions to Decisions.md, newest at top." Friction has to be near-zero. If appending takes more than 90 seconds, it stops happening within two weeks. Miss this and: the architecture decays. Decisions.md freezes in March. By June the AI is operating on a three-month-old map of the business.
The Pitfalls
The Growing-Blob Trap. One ever-expanding memory.md that becomes unsearchable. Fix: split into the four file types on day one and keep each one under 2,000 words by archiving old context to dated files.
The Fresh-Start Trap. Every new project gets a new memory folder. Don't. Fix: one memory system across all projects, with project-specific sub-files referenced from the Context file.
The Over-Structured Trap. Designing the perfect schema before writing the first entry. Fix: start with markdown and headings. Schema emerges from usage. Real templates take three weeks to settle.
The Fabrication Risk. When the AI can't find an answer in memory it sometimes invents one. Fix: the Persona Cascade must explicitly forbid this for operational facts — "if it is not in memory, say not in memory and ask, do not guess."
The Private-Info Leak. Memory files often hold equity, comp, HR, family details. Fix: never load them into a shared or team AI workspace. Local-only or single-user accounts. If a teammate needs access, write a redacted Context-only variant.
The Drill (this week)
Thirty minutes, one folder. Create Memory/ with four files: Identity.md, Context.md, Decisions.md, and an empty People/ subfolder with a .gitkeep. Write one paragraph in Identity — role, business, the three constraints that don't move. Write one paragraph in Context — what's happening this week, top three numbers, biggest open question. Append one entry to Decisions.md for the most meaningful call made in the last 30 days, including the alternatives considered. Now open Claude, paste all three files into a single chat, and ask: "Based on what you just read, what should I be focused on this week, and what is one thing I should stop doing?" If the answer is grounded in the Context file, the architecture works. If it's generic, the Context file is too vague — rewrite and re-test before adding people profiles.
The Tools
| Layer | Primary | Alternates |
|---|---|---|
| File store | Obsidian vault (local markdown) | Plain folder of .md files, Notion database, Mem.ai |
| Load into chat | Claude Projects (knowledge sidebar) | Claude Code CLAUDE.md, Cursor .cursor/rules |
| Cost control | Anthropic prompt caching (cache_control: ephemeral) | None — pay full token cost every call |
| Append automation | Claude Code with bash write access | Manual paste after voice memo, Zapier-to-file |
| Search across memory | Obsidian native search + backlinks | grep, ripgrep, Mem semantic search |
Vendor-neutral on the file store. The opinion: plain markdown in a local folder beats every database for this job because the AI reads it natively and the operator owns the file.
Cross-references
Memory is the Remember surface from The 4-Surface AI Stack — this framework is the implementation detail of that column. The Persona Cascade depends on Memory for its Context Priority Order: the persona file points to memory locations so the AI knows where to look before answering. Forward, two frameworks build on this one: The Operator Vault is the full Obsidian implementation — folder structure, plugins, templates, daily-note flow. The Hand-off Brief Pattern uses Memory entries as the source material a brief references when sending Claude Code on a multi-hour task. Read this framework before either of those; without the four file types in place, both downstream frameworks become harder than they need to be.
One framework. One drill. One week at a time.
The Operator Stack is the architecture. Verala is the practice that runs it on your own communication delivery — voice, pitch, pause, presence. One foundation per week, until it's automatic.
Take the free 5-Foundation Voice Audit → · Book a 30-min intro call →