What is ChatGPT Memory?

A “vanilla” LLM is stateless: it predicts the next token from whatever sits in the current prompt, then forgets when the session ends. ChatGPT’s Memory adds a small, user‑scoped store the assistant can read (to personalize) and write (as you share things worth remembering). In ChatGPT, memory works through two channels: (1) saved memories—facts you explicitly ask it to keep or that it flags as useful—and (2) chat‑history reference, where it can pull relevant details from prior conversations even when you didn’t save them. These persist across sessions and can be toggled independently.

You stay in control. You can review, edit, or wipe items under Settings → Personalization → Manage memories. Turning memory off stops new saves and usage but doesn’t delete existing entries; deleting a chat also doesn’t erase its saved memories—you must remove those directly. If you want a one‑off, clean slate, use Temporary Chat, which won’t reference or update memory and won’t retain that conversation in history.

How LLM Memory works under the hood (conceptually)

Production memory systems distill long chats into compact, structured snippets (think key–value notes or terse summaries), index them semantically, and inject only the most relevant pieces back into the model’s context at answer time. Research and industry patterns add tiers/scopes (profile vs. project), recency weighting or decay/TTL to avoid stale facts, and consent‑first controls. These ideas echo work such as MemGPT, which manages multiple memory tiers to extend usable context, and recent surveys that frame memory along timescales and retention policies. (arXiv)

Why it feels like a helpful colleague. With a curated “you” in context, defaults emerge: ask for “snacks,” and results bias toward your diet; ask for “shirts,” and it starts with your preferred palette—without you restating constraints each time. That reduces prompt bloat and error‑prone repetition while keeping personalization auditable.

Beyond ChatGPT: opening memory to any LLM.

New platforms like mem0 expose memory as a universal layer you can plug into agents built on OpenAI, Anthropic, local models, and more. mem0 provides OSS SDKs and a hosted service; it extracts candidate memories, applies filtering/decay to prevent bloat, semantically retrieves the most relevant items, and injects them at generation time—separate from (but complementary to) document‑centric RAG. For teams that need portability across models or vendors, this “bring‑your‑own‑LLM memory” pattern is compelling. mem0 even reports gains in accuracy and latency from smarter injection and fewer tokens (methodology and numbers in their research notes).

Bottom line: ChatGPT Memory operationalizes long‑term personalization with user controls and narrow, relevant recall; memory layers like mem0 extend the same primitive to any LLM stack. Used thoughtfully—scoped, decayed, and reviewable—memory turns ephemeral chats into durable collaboration.