What is AI memory, and how is it different from a long chat?
AI memory is a persistent store of facts about you that the assistant carries between separate conversations. It is not your chat history. It is a short, model-written profile containing your role, preferences, recurring projects and formatting rules, which the assistant loads into a new conversation before it answers you.
Most people using AI every day have never opened this setting. They keep re-pasting the same brief at the top of every chat because they assume the tool is stateless.
It is not, and has not been for months. Anthropic rolled persistent memory out to all Claude users, free and paid, in March 2026. OpenAI has shipped memory in ChatGPT since 2024 and now runs it as two independent layers.
The distinction that matters: a long chat holds context inside one conversation and dies when you close it. Memory survives the conversation. Getting the second one right is what stops you rewriting your own brief forty times a month.
How does Claude's memory actually work?
Claude builds memory by synthesising your conversations roughly every 24 hours into a running summary, then retrieves from that summary selectively, only when it judges past context relevant to your current request. You control it at Settings, then Capabilities, then Memory, where every stored fact is visible and individually editable.
Two things about this design are worth knowing before you use it.
Retrieval is selective, not automatic. Claude decides per-request whether to pull memory in. That means a fact can be stored correctly and still not appear in an answer, because the model did not consider it relevant. If a memory matters for a specific task, say so in the message.
Synthesis runs on a delay. Because the summary is rebuilt roughly daily, something you mention this morning may not be in memory this afternoon. To make it stick immediately, tell Claude directly in the chat what to remember rather than waiting for it to infer.
On Pro and Max plans, each Project also carries its own separate memory. This is the layer most people miss. A Project for Client A and a Project for Client B do not share memory, which means you can run two conflicting brand voices without them bleeding into each other.
Control options are Pause, which keeps existing memory but stops building new entries, and Reset, which permanently deletes everything and cannot be undone.
How is ChatGPT's memory different from Claude's?
ChatGPT splits memory into two layers that toggle independently. Saved memories is an explicit, user-editable list of facts. Reference chat history is implicit recall of patterns across your past conversations. Both live at Settings, then Personalization, then Manage memories, where you can view, edit or delete individual entries or clear the whole list.
The practical difference from Claude is transparency versus reach.
ChatGPT's saved memories list is explicit and auditable, but reference chat history operates in the background and you cannot inspect what it drew on for any given answer.
Claude shows you a single reviewable list, so it is easier to trust and easier to correct, but it retrieves less eagerly, so you sometimes have to prompt it to use what it knows.
Neither is better. They fail differently. ChatGPT is more likely to carry over something you wanted left behind. Claude is more likely to ignore something you wanted carried over.
One control worth setting today: in ChatGPT, the memory toggle and the data-training toggle are separate switches. Turning memory off does not change your training-data setting, and turning training off does not stop memory being written.
What should you actually put in AI memory?
Put in facts that are stable, reusable across many tasks, and expensive to retype. Your role, your audience, your output formats, your hard constraints and your banned words. Do not store one-off task details, anything time-bound, or client information you would not want surfacing in an unrelated chat.
The fastest way to set this up is a single seeding message. Paste this into a fresh conversation in Claude or ChatGPT, edit the values, and send it once.
Try this prompt:
--- Remember the following about how I work, and apply it by default in future conversations unless I say otherwise.
--- ROLE: I am a marketing manager at a 40-person B2B software company in Hong Kong. I write for operations and finance decision-makers, not for engineers.
--- OUTPUT DEFAULTS: British English spelling. Short paragraphs, two to three lines each. No bullet points unless I ask. Never open with a throat-clearing sentence about how important something is.
--- HARD CONSTRAINTS: Never invent statistics. If you do not have a real source, say so and use directional language instead. Always give me the shortest version that still answers the question.
--- BANNED WORDS: leverage, seamless, robust, game-changing, in today's fast-paced world.
--- WHEN I ASK FOR A DRAFT: give me one version, not three, and add a one-line note on what you were unsure about.
--- Confirm what you have stored, in a list, so I can correct it now.
That last line is the part most people skip. Asking the model to read its own memory back to you turns a guess into a verified record, and it takes ten seconds.
Where does AI memory break, and how do you fix it?
Memory breaks in four predictable ways: stale facts that were true six months ago, contradictions between entries the model resolves arbitrarily, context bleed between unrelated clients, and privacy leakage into chats you did not expect. All four are fixable by editing the stored list rather than by prompting harder.
Stale facts. A memory entry saying you work at a company you left in March will quietly shape every answer. Nothing flags it. Open the memory list monthly and delete anything no longer true.
Contradictions. If one entry says "always use British English" and a later one says "match the client's US style guide", the model picks one and you will not be told which. Write constraints as conditionals instead: "British English by default; US spelling only when I say the client is US-based."
Context bleed. This is the strongest argument for Claude's project-scoped memory on Pro and Max. If you handle multiple clients in one account and you are not using separate Projects, expect one client's tone to appear in another's draft eventually.
Privacy. Do not put client names, unreleased figures, salary data or anything under NDA into a global memory that any future chat can read. Use a Project, or keep it in the conversation only.
There is also an honest limitation worth naming. Memory improves consistency, but it does not fix a weak prompt. If your instructions are vague, memory just makes the model vaguely wrong more consistently. Memory is a floor for quality, not a ceiling. For the discipline of assembling the right context per task, see our guide to context engineering.
How do you audit your AI memory in 20 minutes?
Run this once and you will not need to touch it again for a month. Open your memory list, delete what is stale, seed what is missing, then test with a cold prompt that deliberately gives the model no context and see whether the output already matches your standards.
--- Minute 0 to 5: Open Claude at Settings, Capabilities, Memory, and ChatGPT at Settings, Personalization, Manage memories. Read every entry. Delete anything untrue, time-bound or embarrassing.
--- Minute 5 to 10: Send the seeding prompt above in each tool. Read back the confirmation list and correct anything wrong.
--- Minute 10 to 15: Open a brand-new conversation and send one deliberately bare instruction, for example "Draft the opening paragraph for a client update about a two-week delay." Give it nothing else.
--- Minute 15 to 20: Compare the output to your standards. Every gap you find is a missing memory entry. Add it, and re-run the same bare prompt.
The cold-prompt test is the whole point. If a naked instruction produces something you would send, memory is configured. If it produces generic filler, memory is not doing its job yet and you now know exactly which fact is missing.
Claude memory and ChatGPT memory at a glance
Claude
--- Availability: all users, free and paid, since March 2026
--- Settings path: Settings, Capabilities, Memory
--- How it builds: model-written summary synthesised roughly every 24 hours
--- Retrieval: selective, only when Claude judges past context relevant
--- Scoping: separate per-Project memory on Pro and Max plans
--- Controls: view and edit individual entries; Pause; Reset (permanent)
ChatGPT
--- Layers: saved memories (explicit, editable list) and reference chat history (implicit)
--- Settings path: Settings, Personalization, Manage memories
--- Retrieval: both layers active by default, toggled independently
--- Controls: view, edit or delete individual entries; Clear all; toggle memory off
--- Note: the memory toggle and the data-training toggle are separate settings
The takeaway
Twenty minutes of memory configuration removes a task you were doing several times a day without noticing. That is the entire return, and it is larger than it sounds, because the cost of re-briefing an assistant is paid in attention rather than in time.
The reason this setting sits unopened is not difficulty. It is that nobody told you it was there. Good tools should not need a treasure hunt, and good technology should meet you where you already are.
We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Turn a Configured Assistant Into a Working System
Memory makes one assistant consistent. Turning that into a workflow your whole team can run is the next step, and we'll walk you through every step, from tool setup and memory design to deployment and handover.
Written and reviewed by the UD AI team, Hong Kong. Last updated 31 July 2026.
