Short-term context vs. long-term memory

Every AI conversation has two kinds of recall. Short-term context is the running transcript of the current chat—it disappears when the session ends. Long-term memory is different: selected facts are extracted from conversations, stored separately, and re-inserted into the prompt next time you talk.

• Short-term context: full recent messages; high fidelity; resets per session.
• Long-term memory: compact saved facts; durable across sessions; editable by design.
• Without long-term memory, a companion restarts from zero every time you return.

How a memory layer extracts and reuses details

After or during a conversation turn, the system checks whether anything durable was shared—a preference, a personal detail, an ongoing story. If yes, it writes a short structured note such as “prefers calm evenings to crowded plans.” In later sessions those notes are added to the prompt alongside the character definition, which is why the companion can follow up on last week's photo walk.

Why user control is non-negotiable

Memory that users cannot inspect becomes surveillance. The minimum controls any memory-based platform should offer:

• Visibility: every stored memory readable by the user inside the chat interface.
• Editing: fix a wrong detail instead of letting one misunderstanding shape every future reply.
• Selective deletion: remove one memory without wiping the whole relationship.
• Bulk erasure: clear all memories in one action.
• Opt-out: switch memory off entirely if you prefer stateless chats.

What responsible platforms do NOT store

A memory system built for companionship does not need payment details, precise location, identity documents, or contact lists—and it should never use your intimate conversations as advertising fuel. If the business model depends on profiling you rather than charging transparently for the product, treat the memory feature as a liability.

Questions to ask before committing to a memory-based companion

Five questions separate the transparent platforms from the black boxes:

• Can I read every memory right now, in the product, not just in a policy document?
• Can I edit or delete individual memories myself, without contacting support?
• Is there a single button that erases everything?
• Does the character ever claim to be human? (It should be technically prevented.)
• What happens to my memories if I stop paying or delete my account?