EPS Digital case study / an ai that remembers my business

AI Brain case study

An AI that remembers my business, so I never start from zero

My AI assistants read a private folder of plain files before every session. They start knowing my rules, my projects and my past mistakes.

read by
claude codecodex
started 2026-06-19
Typical AI chat Typical AI chat. New chat starts from zero, you re-explain who you are and your rules, lessons are forgotten, the same mistake happens again.New chatstarts from zeroYou re-explain everythingwho you are, rules, historyLessons are forgottenlast mistake is lostSame mistake againback to step one
Every chat starts from zero.
source: how chat assistants work without saved context
My AI, with the vault My AI with the vault. New session starts by reading files, reads rules, context and hot cache, knows past decisions and lessons, ready to work without re-explaining.New sessionstarts by reading filesReads the vaultrules, context, hot cacheKnows the businesspast decisions and lessonsReady to workno re-explaining
Every session starts knowing the business.
source: my vault setup, rules file read first each session

Why I built it

I was tired of starting over

The frustration

  • Re-explaining who I am
  • Repeating the same mistake
  • Losing past decisions

What the vault holds

268 files
  • rules
  • who I am
  • hot cache
  • session log
  • project notes
  • wiki
  • lessons
plain markdown, as of 2026-10-04

What I get

An AI that starts informed

  • No re-explaining
  • Mistakes fixed once
  • One memory across projects
source: my own use since 2026-06-19

01 / Problem and cost

Forgetful AI repeats its mistakes

Problem

Every chat starts from zero

You re-explain who you are, your brand rules and past decisions.

Cost

The same mistake, again

Lessons from one mistake are forgotten, so it repeats.

02 / How it works

A folder of plain files, read in order

The AI reads six layers at the start of every session, top to bottom.

Vault layers, in read order Vault layers in read order: rules file, who I am, hot cache, session log, project folders with a locked private folder, wiki. Rules filehow to work, preferences, model tiering Who I amidentity and ongoing facts Hot cache~500-word fast summary, refreshed per source Session logcross-project, dated Project folderseach: own context + dated memory log EPS Digital ESTableTennis site Instagram the two apps private finance: locked, never reused Wiki knowledge basebuilt from source documents 1 2 3 4 5 6
Six layers, read top to bottom. The private folder stays locked.
source: my vault rules file, read order
Adding a source Ingest flow: source, summary page, 5 to 15 updated pages, then index, hot cache and log. A lint loop checks contradictions, stale claims and orphan pages. Source documentdropped in raw/, never edited Summary pageone per source 5 to 15 pages updatedrelated people, ideas, projects Index Hot cache Log Lint health checkcontradictions, stale claims,orphan pages repeats on a schedule
One source touches 5 to 15 pages. Lint catches drift.
source: my wiki workflow rules

03 / Human role

The AI proposes. I approve.

I approve core changes

Rules and context files change only when I say yes.

Sources stay untouched

Raw documents are never modified.

Finance stays walled off

Its figures never get reused elsewhere.

No logins in notes

Memory notes never store logins or keys.

source: my vault rules file

04 / Results

What it holds today

Counts as of 2026-10-04.

88 wiki pages by type 88 wiki pages by type: 36 source summaries, 33 concepts, 12 entities, 7 synthesis pages.source summaries36concepts33entities12synthesis and Q&A7
Wiki pages by type, 88 in total
TypePages
source summaries36
concepts33
entities12
synthesis and Q&A7
source: my vault file count, 2026-10-04
268

markdown files in the vault

excludes the private folder
82

saved memory facts

one fact per file
44

logged wiki operations

since 2026-06-19
6

production pipelines

5 live, 1 retired

Each pipeline is a saved workflow the AI loads when the job matches. Each records its known failure modes. The long-form video one documents 40. Pipelines propose their own improvements, and I approve before anything changes.

Learn once

A mistake gets fixed once

Mistakesomething goes wrong
Saved lessonone note in the vault
Automatic checkapplied next time

Wrong folder

A hosting mix-up put files in the wrong website folder. Every deploy now checks the folder first.

Brand rules

"No em dashes" and the exact ebook title are saved once. Every piece follows them.

Changed affiliate tag

I recorded the new tag. Every new link uses it.

source: saved memory notes, real examples

Cost control and automation

Cheap models do most of the work

10-80-10 model tiering 10-80-10 model tiering: plan 10 percent on the strongest model, build 80 percent on cheaper models, review 10 percent on the strongest model.108010Planstrongest modelBuildcheaper modelsReviewstrongest model
Plan and review on the strongest model. Build on cheaper ones.
source: tiering rule written into my vault rules
  1. Vault backs up to cloud storage
  2. Downloads folder tidied
  3. App audit runs
source: my scheduled jobs on this machine

05 / What it means for your business

Your business, already known

This vault runs EPS Digital itself, and the whole ESTableTennis content operation. The same pattern fits yours.

Your SOPs

Your client history

Your brand rules

Your lessons

An AI that already knows them

What to insist on

  • Plain files you own and can read
  • The AI proposes, you approve
  • Sensitive data walled off
  • Automatic daily backup

How I measured this

Where each number comes from

  • 88 pages (12 / 33 / 36 / 7), 268 files, 82 facts, 44 operations: counts of files in my vault as of 2026-10-04. The 268 excludes the private finance folder. The 44 is entries in the wiki operation log since 2026-06-19.
  • 6 pipelines, 40 failure modes: saved workflows for ESTableTennis production. Five are live and one is retired. The 40 is the failure-mode list in the long-form video pipeline.
  • 10-80-10: a rule written into my vault rules file, not a measured saving.
  • 6:00 and 7:00, monthly: my scheduled jobs on this machine.
  • I claim no time-saved number for the vault itself. These are counts of what exists, not results from a controlled test.

Your business

Give your AI a memory

On a free AI Fit Call we look at what your AI should already know, and what it should never touch.

Book a free AI Fit Call

One conversation. A clear view of where to start.