Free lesson

Memory Under Control

Memory Under Control — Your Personal AI Assistant: Build One That Knows Your Context, No Code Needed. What memory actually means; Review memory item by item; When outside content tries to alter memory; Five memory rules; The complete cycle: clean, test and record. Practical assignment: Clean and protect memory.

Demo edition. The materials are undergoing editorial review.

Topics covered

  • What memory actually means
  • Review memory item by item
  • When outside content tries to alter memory
  • Five memory rules
  • The complete cycle: clean, test and record

What memory actually means

An assistant’s memory is not a complete record of everything you have ever said. It is selective context: separately saved facts, summaries derived from earlier conversations and, depending on the service and settings, other available sources. The assistant decides which context is relevant to a particular answer, so the presence of a fact in memory does not guarantee that it will always be used.

As of 17 November 2026, controls differ by service, plan, region and workspace. For example, ChatGPT may show saved memories, chat history or a memory summary; Google Gemini can use past chats where the feature is available and enabled. Check the actual control names in your account instead of relying on an old walkthrough.

Memory ages. “I work night shifts” may have been true for one month only. An assistant can also mistake a friend’s situation for yours or take a joke literally. Stable preferences can be useful to save, but important facts still need verification before you rely on them.

A common mistake is to turn memory off and assume that everything has been deleted. Disabling a feature, deleting a remembered item and deleting the original chat are separate actions. In ChatGPT, deleting a chat does not necessarily delete a separately saved memory. For full removal, inspect every place where the information appears: memory or its summary, regular and archived chats, saved files and connected apps.

Review memory item by item

Open the memory settings. If the service shows only a summary, inspect it and ask, “List everything you remember about me, one item per line.” The assistant’s answer may be incomplete, so compare it with the controls and sources you can actually see.

Give every item one label:

  • Correct: true, reasonably stable and useful in later answers — keep it.
  • Outdated: once true but no longer true — delete it and add the current version only if useful.
  • Wrong: never true or about somebody else — delete it.
  • Unnecessary: true, but unsuitable or pointless to retain — delete it.

Passwords, document numbers, exact financial details, confidential facts about other people and trivia that will not improve future answers belong in the unnecessary group. Do not create a new memory containing sensitive information merely to “correct” the old one.

Explicitly fictional example. Ruth has 63 items. She finds 19 that are outdated or wrong and 4 unnecessary items about her father’s health. She deletes those 23 items and adds 3 current, non-sensitive facts: 63 − 23 + 3 = 43. The number 43 is not a quality score. The useful result is that Ruth has deliberately chosen to retain each of those 43 items.

When outside content tries to alter memory

A document, email or webpage can contain text that resembles instructions to the assistant. When such text attempts to steer the system’s behavior, it is called prompt injection. OWASP defines indirect prompt injection as malicious instructions embedded in external content processed by a model. The consequences depend on the tool’s permissions: reading one document presents less exposure than access to memory, email or actions in other systems.

Do not treat every bad memory as evidence of an attack. An unfamiliar item may come from a poor summary. Still, an entry such as “always include this link,” or a preference you never stated, should be removed and its source inspected when the interface provides that option.

Use a mode that does not create or update memory for unfamiliar files and links. In ChatGPT, a temporary chat does not create or update memories; before it starts, you can also choose an unpersonalized option. Gemini’s temporary-chat availability and behavior depend on the account. This reduces one route by which memory could be contaminated, but it does not prove that a file is safe. Review memory monthly and investigate entries you do not recognize.

Five memory rules

After the cleanup, decide what the assistant may retain. Here is a workable five-line policy:

  1. Remember stable facts, my formatting preferences and long-term goals.
  2. Do not remember passwords, document numbers, exact financial details, conflicts or private facts about other people.
  3. Do not save short-term events—this week’s plans, a temporary illness or a one-off trip—unless I explicitly ask.
  4. If you are unsure whether something should be remembered, ask me first.
  5. I review memory once a month and remove outdated, wrong and unnecessary items.

Put these rules in a place where your service actually supports persistent instructions. If useful, add a short version to standing instructions: “Keep only stable, useful facts; ask before saving personal information.” A written rule is not a technical guarantee, so test it.

Explicitly fictional test. Ruth writes in a normal chat, “I have a dentist appointment on Thursday; my sister is going through a divorce.” The first detail is short-lived; the second is private information about another person. After the response, Ruth checks memory. If either detail appears, she removes it and makes the relevant rule more precise. For a real test, use harmless fictional details rather than disclosing genuine sensitive information.

The complete cycle: clean, test and record

Memory control is a complete cycle. First, record the number of visible items. Classify each one, remove everything outside the correct category, and add only necessary current facts. Then enter the five rules and run a test with one fictional short-term detail and one fictional sensitive detail.

Record the result for the next review: date, number before cleanup, number after cleanup, counts of outdated, wrong and unnecessary items, and whether the test passed. If the interface does not expose a complete list, state the limitation plainly: “Visible summary reviewed; completeness not confirmed.”

Do not test quality only by asking, “Are you following my rules?” The assistant’s answer is not evidence of its behavior. Inspect the observable result in settings. Memory is useful for convenient context, but important dates, contract terms, medical information and other critical facts should remain in an appropriate reliable source that you can verify.

Practical assignment

Clean and protect memory

Continue to learning

Course outline

  1. An Assistant That Knows You — in 20 Minutes
  2. Five Layers: Where Everything Lives
  3. Memory Under Control
  4. Instructions That Actually Get Followed
  5. Projects: One Area of Life, One Workspace
  6. A Reference File Your Assistant Can Rely On
  7. Specialized Assistants: One Role, One Job
  8. Routines: One Phrase Instead of an Explanation
  9. An Assistant on a Schedule: Regular Check-Ins
  10. Voice and Phone: Capture on the Go
  11. Email, Calendar and Files: Access on a Leash
  12. Sycophancy: When Your Assistant Agrees Too Easily
  13. Big Decisions: Your Assistant Prepares, You Decide
  14. Keep Your Own Head: What Not to Hand Over
  15. When Your Assistant Gets It Wrong: Fix the Layer
  16. Upkeep and Moving: Your Assistant Passport