Inside AI memory: what apps really save about you and why it matters

"Memory" sounds like the AI is quietly getting to know you. The reality is more mundane and more important: the app is writing notes about you and pasting them into every future request. Here's what that actually looks like.

Memory is a file, not a mind

Models are stateless. Each call starts fresh. What vendors call memory is a store — often a short list of facts, preferences, or summaries — that the app injects into the context stack before the model sees your prompt.

What usually gets written

• Stated preferences ("I prefer concise answers," "I work in healthcare").
• Inferred facts (your role, your projects, your tone).
• Summarized past conversations that the app thinks might matter later.

Where memory lives

Sometimes in a plain JSON blob attached to your profile. Sometimes in a vector database so the app can retrieve the relevant memories per question instead of injecting all of them. The second approach scales better and leaks less.

The three risks leaders should name

Privacy drift. Memory written in one context (a casual chat) resurfaces in another (a formal report). Users rarely realize.

Stale truth. Yesterday's "I live in Boston" becomes today's wrong assumption. Memory needs expiration policies.

Prompt injection. A memory populated from untrusted text (an email, a document) can carry instructions that hijack later sessions.

Governance questions to ask any vendor

• Can users view, edit, and delete their memory?
• What's the retention policy, and who can audit writes?
• Is memory partitioned by project, role, or data classification?
• How are memories vetted before being used as instructions?

For an organizational posture on these questions, see our Responsible AI-by-Design Framework and Fractional Chief AI Officer services.