I have the cure for amnesia

Hi {{first_name | Reader}},

I've always been fascinated by how the mind works.

I mean the practical kind. What happens in the seconds before I walk into a client meeting and ask the right question? How does the right context surface at the right moment? Why do I remember a detail from six months ago that changes the entire conversation?

For years I took it for granted. Pattern matching. Experience. Gut feel.

It was not until I started automating myself that I had to make the implicit explicit. Building an AI that does my job forced me to answer questions I had never asked: What do I actually retrieve before a meeting? What do I always carry in my head? What do I only look up when I need it?

Last week I promised you three companies and one truth. That is coming. But this week I built something that changes how every future send works, and I need to show you before we go further.

Automating yourself is the most honest mirror you will ever build. And the first thing the mirror showed me was that my mind runs on two systems.

Your brain has two memory systems and they serve different purposes.

The hippocampus handles short-term memory. Recent, contextual, lossy. It holds what happened today, what you read this morning, the client name you just looked up. It is fast, it is fragile, and it overwrites itself constantly.

The neocortex handles long-term memory. Consolidated, permanent, retrieved on demand. It stores the patterns you have accumulated over years. The benchmarks you know by heart. The frameworks you reach for without thinking. It is slow to build and hard to overwrite.

Two systems. Two purposes. One brain.

So I built both.

Short-term memory: the Brief.

The Brief is the document my AI reads before every conversation, every chat session, every report. It holds the current state of a client: where the numbers stand, what changed this week, which flags are active, what the last meeting covered. It is recency-weighted, compact enough to fit a small context window, and updated every time the system learns something new.

Every previous version of the Brief is still there. Snapshot history. You can scroll back and watch how the story of a business changed over six months. You can catch the moment the AI's understanding diverged from reality. Versioned thinking.

Long-term memory: the knowledge base.

The knowledge base holds everything the system has ever learned, organized by domain. Benchmarks. Frameworks. Client patterns. Tax strategies. Consulting methodologies extracted from 14 months of practice across 180 meetings with real clients.

It does not load into every conversation. That would flood the context window with irrelevant information. Instead, it sits on the shelf. When the AI needs it, it reaches for it. Semantic search, not keyword search. Meaning, not strings.

This is the top shelf. Always loaded. 7 context types filled, 35,000 tokens. Green means fresh. Amber means aging. Gray means blind spot.

And this is the bottom shelf. Searchable on demand. 27 documents, 63,700 tokens. The AI retrieves them only when the question requires it.

Two memory systems. Two retrieval paths. Two staleness rules. The Brief is for the next conversation. The knowledge base is for the next quarter.

The memory does not build itself. There is a pipeline behind it, and it starts where all knowledge starts: in communication.

A CFO's real expertise does not live in the general ledger. It lives in the email thread where the client explained why revenue dipped. The meeting where the bookkeeper admitted the accrual was wrong. The quick message about a delayed invoice. The correction. The nuance. The "oh, by the way" that changes how you read the whole quarter.

No report captures this. No spreadsheet stores it. Every decision depends on it.

So I built a pipeline that captures it:

1. Data sources. Transcripts from every client meeting (via Fireflies). Emails synced daily from Gmail. Documents uploaded from Drive or dropped into the inbox.

2. Ingestion. Raw content enters the document inbox. An auto-classifier assigns a type, a period, a confidence level. High-confidence items get auto-approved. Everything else waits.

3. Extraction. Claude Opus reads each transcript and pulls knowledge nuggets: structured insights tagged by type (trend, risk, pattern, preference, anomaly) and category (cash flow, revenue, ops, relationship, compliance). Signals, not summaries. The specific kind of financial nuance that changes a report.

4. The human gate. Every nugget stops here. Approve. Reject. Edit the metadata. Nothing enters the permanent database without a human saying yes. Nothing.

5. Permanent storage. Approved nuggets enter the knowledge base. Optionally promoted into the Brief if the human decides this insight should always be in context.

6. Retrieval. The AI queries the knowledge base through semantic search in chat, in briefs, in pre-meeting prep. The Brief loads automatically. The rest loads on demand.

Here is what the extraction produced this week.

Three nuggets from one week of client work. Each one tagged by domain. Each one waiting for the human to approve.

The pipeline runs continuously. Transcripts arrive after every meeting. Emails sync daily. The extraction cron finishes at 4am, and by the time I open the inbox at seven, the nuggets are waiting. The human reviews. The memory grows.

Last week I told you clean data was the floor. Context is the next floor up. And the gatekeeper is the person holding the key to the staircase.

This is the professional responsibility nobody talks about when they talk about AI in finance.

Not "did the AI get the right answer" but "did the AI have the right context?"

Polluted context is worse than no context. A hallucination is obvious. The numbers do not tie. You catch it in five seconds. But an AI that retrieves a stale insight from six months ago, or misses a recent conversation where the client changed direction, will produce a confident, well-formatted, completely wrong report.

And the CFO signs it.

The gatekeeper's job has three parts:

- What enters context. Approve or reject at the inbox. Every nugget, every document, every insight.

- How it is retrieved. Semantic search relevance. Staleness thresholds. Confidence weighting.

- Whether it is being retrieved at all. Blind spot detection. The system flags context types that have no recent data.

The blind spots are the real risk. Omission, not hallucination.

One checkbox. The human decides what the AI remembers permanently. Approve, and the nugget enters the knowledge base. Check "Promote to context," and it enters the Brief. Every future conversation with this client starts from a more informed place.

This is not a feature. This is the job.

Not all knowledge is the same. The system separates it into three scopes, each with different rules.

Client-specific. This client's nuances. The bookkeeper who always posts late. The CEO who wants the P&L before the cash flow. The revenue recognition decision from Q2 that affects every forecast. Stale after 90 days without a new signal.

Company-wide. Cross-client patterns and benchmarks. What "good" bookkeeping quality looks like across a dozen clients. How SaaS margins differ from manufacturing. Patterns that repeat. Refreshed quarterly.

Meta-learner. General CFO knowledge extracted from 14 months of practice. Not tied to any client. Frameworks, benchmarks, vocabulary, decision patterns. 42 articles across 12 domains. Evergreen.

Each scope has different staleness rules. Client data older than 90 days gets a yellow badge. Then red. The meta-learner articles stay green forever. The system knows the difference because I told it the difference.

You have seen the architecture. Two types of memory. A pipeline from raw communication to structured intelligence. A human gatekeeper. Three knowledge scopes with different staleness rules.

What you have not seen is what I tell the AI to look for.

The extraction prompt that catches financial nuances a general-purpose LLM misses. Not "summarize." The specific signal pattern. The metadata schema that makes every nugget searchable, scorable, and expirable. The gatekeeper rules expressed as CLAUDE.md rules you can drop into your own agent. The recipe that decides what always loads and what stays on the shelf. And the staleness policy that keeps the whole system honest.

Behind the wall, five system design decisions you can copy without writing a single line of code:

1. The extraction prompt. What I tell Claude Opus to look for in a meeting transcript. Not "summarize." The specific signal extraction pattern that catches the passing comment about switching bookkeeping software that a cheaper model misses.

2. The metadata schema. What fields a knowledge nugget has and why. Copy-pasteable for anyone building their own system with Claude, ChatGPT, or any LLM.

3. The gatekeeper rules. Seven CLAUDE.md rules that govern what auto-approves, what requires human review, and what gets rejected. Drop them into your own agent file.

4. The context assembly recipe. What always loads. What stays on the shelf. And the one question that decides which category something falls into.

5. The staleness policy. How old is too old. What triggers a refresh. What triggers a badge. A table you can adapt for your own practice.

None of this requires you to know React or Supabase or TypeScript. It is operational architecture. The same kind of thing a fractional CFO would document in a playbook. Except this playbook teaches an AI to remember your business.

€19 per month. Cancel anytime. One year for the price of ten months.

The story stays free. Forever. The principle stays free. Forever. What moves behind the wall is the how. The prompts, the rules, the schema, the exact decisions you can copy into your own practice.

Reply and tell me: what does your AI know about your business? I read every one. Most people answer "nothing." That is the blind spot.

Or take the Scalable Finance Wheel if you want a broader diagnostic on the finance function.

See where your finance function stands →

Next week: three companies, one truth. I run the scanner across three related entities in two countries. Same chart of accounts, different books, different stories. Saturday.

Samer