The Curriculum / Reader / 06 — Memory Hygiene
LEVEL 1 · ESSENTIALS · INDIVIDUAL TRACK

06 — Memory Hygiene

This page compiles 3 files from the repository, verbatim, in reading order. The living version: this folder on GitHub.

level-1-essentials/individual/06-memory-hygiene/README.md

06 — Memory Hygiene

Every major model now has memory. Curated memory is a superpower. Un-curated memory is drift.

The rule

Memory should contain: - Durable facts (name, role, projects, timezone) - Persistent preferences (voice, formatting, length) - Standing constraints (privacy line, banned topics)

Memory should NOT contain: - Ephemeral tasks - Personal secrets - One-off preferences ("make this shorter") - Anything that changes weekly

Files in this folder

  1. what-to-tell-each-model.md — the initial memory setup script for every provider
  2. memory-audit-template.md — monthly review template

Time budget

Definition of done

level-1-essentials/individual/06-memory-hygiene/memory-audit-template.md

Memory Audit Template

Run this monthly. 10 minutes.

Step 1 — Enumerate what each model remembers

In each model, ask:

"List everything you remember about me, verbatim."

Copy the output into this file, one section per model.

ChatGPT memory dump — [DATE]

[paste]

Claude Project instructions — [DATE]

[paste]

Perplexity AI profile — [DATE]

[paste]

Gemini personal context — [DATE]

[paste]

Step 2 — Score each memory item

For every stored fact, decide: - ✅ Keep — still true and useful - ✏️ Update — needs revision - 🗑️ Delete — outdated, ephemeral, or shouldn't have been stored

Step 3 — Common corruption patterns to look for

Step 4 — Prune and update

Step 5 — Log the audit

Record in this file:

Date Model Entries reviewed Deleted Updated Notes

Step 6 — Re-run the calibration test

Same test as initial setup:

"Given what you know about me, draft a 3-sentence Slack message declining a lunch invitation from a client."

If any model's output drifted, memory needs more pruning.

level-1-essentials/individual/06-memory-hygiene/what-to-tell-each-model.md

What to Tell Each Model

Use this script once for each provider that supports memory. Copy the block, paste it as a message with the phrase "Please remember the following about me:"

The universal memory block

Please remember the following about me:

Identity:
- My name is [NAME]. I go by [PREFERRED].
- Based in [CITY], [TIMEZONE].
- I work as [ROLE] at [COMPANY / SOLO].

Projects (active):
- [PROJECT 1] — [one line]
- [PROJECT 2] — [one line]
- [PROJECT 3] — [one line]

People I mention often:
- [NAME] — [role, relationship]
- [NAME] — [role, relationship]

Voice preferences:
- Direct over diplomatic.
- Match my length.
- No AI tells (delve, certainly, in the ever-evolving landscape).
- Give recommendations, not option menus.
- Disagree plainly when I'm wrong.

Formatting defaults:
- Prose over bullets when content flows.
- H2 headers in anything over 300 words.
- No emojis unless I use them first.
- Tables only for comparisons.

Constraints:
- Never process credentials, keys, SSNs.
- Warn if I paste secrets.

When I ask for a decision or recommendation, pick one and justify. Do not list five options with hedged pros and cons.

Per-provider setup

ChatGPT (Plus/Pro)

  1. Settings → Personalization → Memory → toggle ON
  2. Open a new chat. Paste the universal memory block prefixed with "Please remember the following about me:"
  3. After it acknowledges, ask: "What do you know about me?" — verify accuracy.
  4. Also fill Custom Instructions:
  5. "What would you like ChatGPT to know about you?" — paste identity + projects
  6. "How would you like ChatGPT to respond?" — paste voice + formatting + constraints

Claude (Pro/Max)

  1. Claude does not have global memory like ChatGPT. Instead, use Projects.
  2. Create a Project called "Personal — Default"
  3. In Custom Instructions, paste the entire universal memory block
  4. Upload your personal-style-guide.md as a knowledge file
  5. Use this project as your default chat home

Perplexity

  1. Settings → Preferences → AI Profile
  2. Fill "Introduce yourself" — paste identity + projects
  3. Fill "Your interests" — bullet your active projects and expertise areas
  4. Enable web memory (opt-in per your privacy line)
  5. For projects that need persistent context, create Spaces per project

Gemini Advanced

  1. Settings → Personal context → toggle ON
  2. Extensions → connect Google Workspace (Drive, Gmail, Docs, Calendar)
  3. Create a Gem called "Adam Default" with the universal memory block as its instructions
  4. Set that Gem as your default in Gemini app

Microsoft Copilot

  1. Copilot uses Microsoft Graph for context — no separate memory input required
  2. Connect Outlook, OneDrive, SharePoint if using enterprise
  3. For consumer Copilot, use the "Personalize" settings

Grok

  1. Personalize → Custom Instructions → paste the compact about-me block (see 03-system-prompts/about-me-block.md)

Test it

After setup, in each model, ask:

"Given what you know about me, draft a 3-sentence Slack message declining a lunch invitation from a client."

If the tone and length feel right in all four models, memory is calibrated.

← 05 — Model Routing 07 — Automation →