# AI Memory, Explained > How ChatGPT, Claude and Gemini remember you, what memory layers like Mem0 and Zep sell to developers, where AI memory breaks, and the five tests any memory you rely on should pass AI memory became its own category in 2026: every major assistant now remembers by default, and a funded wave of memory-layer startups sells it to agent builders. This mindspace maps the four layers of AI memory, the products in each market, the failure modes nobody shows you (invisible, stale, locked in, unsourced), and what memory you can actually see and own looks like. _Category: ai · Tags: #ai-memory #llm-memory #chatgpt-memory #claude-memory #gemini-memory #ai-memory-layer #persistent-memory #mcp-memory #context-window #rag #knowledge-graph #visible-ai-memory · Published: 2026-09-26_ ## How AI Memory Works ### The Model Remembers Nothing A large language model is **stateless**. Its weights are fixed after training, and every reply is computed from one thing only: the text in front of it right now. Close the chat and nothing of it is left in the model. So every form of AI memory is the same trick: **decide what to put back in front of the model next time.** The layers below are four answers to that question, each with a different trade-off between what it keeps, what it loses, and who can see it. ### Layer 1: The Context Window The context window is the model's **working memory**: everything in the current conversation, up to a token limit. Windows have grown from a few thousand tokens to hundreds of thousands or more, which makes a single chat feel like it remembers. But a long window is not memory. It ends when the chat ends, costs more with every token, and the model attends less reliably to details buried in the middle. **Long context postpones forgetting; it does not prevent it.** ### Layer 2: Saved Memories Saved memories are short facts the assistant **extracts about you** and injects into future chats: your name, your job, that you prefer short answers, that you are training for a marathon. This is what most people mean by "ChatGPT memory" or "Claude memory". It is cheap and it personalizes well. Its weakness is that the assistant, not you, decides what is worth keeping and how to phrase it, and a one-line summary of you loses the reasoning behind it. ### Layer 3: Retrieval (RAG) Retrieval-augmented generation keeps the raw material (past chats, documents, notes) in a searchable store and **fetches the most relevant pieces** into the context window when a new question arrives. Referencing your past chat history works this way. Retrieval scales to huge archives, but it finds passages by similarity, not by meaning. It can surface a stale draft next to the final decision and cannot tell which one you settled on. **Search is recall, not understanding.** ### Layer 4: Structured Memory Structured memory stores knowledge as **entities and relationships**: this decision depends on that assumption, this evidence challenges that claim. Knowledge graphs, temporal graphs that track how a fact changed over time, and visual thought maps all live here. Structure is what lets memory answer *why*, not just *what*. It is also the only layer a human can realistically **read at a glance**, because relationships can be drawn, not just listed. ## Who Sells AI Memory ### AI Memory Became a Category Between early 2024 and early 2026, the three largest assistants went from stateless chatbots to systems that **keep long-term personal context by default**. At the same time, a set of startups raised venture money to sell memory as infrastructure to anyone building an agent. That gives the category two very different markets: **built-in memory for people**, which you use without seeing, and **memory layers for developers**, which you wire into software. Knowing which one a product belongs to is the first step in comparing them. ### ChatGPT Memory OpenAI began rolling out memory in ChatGPT in 2024: **saved memories** the assistant writes when you tell it something worth keeping, later joined by the ability to **reference your past chat history**. It behaves like a personal assistant's memory: who you are, how you like answers, what you are working on. You can review and delete saved memories in settings, but what the model draws from your chat history is not laid out as a list you can edit. *Check OpenAI's current help pages: the controls change often.* ### Claude Memory Anthropic rolled out Claude's memory to Team and Enterprise plans in **September 2025**, to Pro and Max in October, and to the free tier in **March 2026**. Its launch emphasized memory you can **read, edit and delete**, scoped to projects so work contexts stay separate, and it added an import tool that pulls context from ChatGPT and Gemini exports. For developers, Claude Code keeps project memory in plain files (`CLAUDE.md`) that live in the repository, which is memory as a document rather than a hidden profile. ### Gemini Memory Google's Gemini treats memory as part of your **Google account** rather than a separate store: saved preferences plus personal context drawn from the Google apps you allow it to use. That makes it strong inside the Google ecosystem and weaker as a record you carry anywhere else. Google also shipped a way to import context from other assistants, so all three vendors now compete on **bringing your memory in**. None of them is built to send it back out. ### Memory Layers for Agent Builders For developers, memory is sold as a layer under the agent. **Mem0** extracts and retrieves memories for personalization. **Zep** (on its open-source Graphiti engine) keeps a temporal knowledge graph of how facts change. **Letta**, from the MemGPT research, splits memory into core, recall and archival tiers. **Supermemory** and **Cognee** round out the field. These are excellent infrastructure, and they answer a different question: *how should my app remember its users?* Not: *how do I keep and see my own thinking?* ## Where AI Memory Breaks ### The Memory You Can't See The core problem with most AI memory is that it is **invisible by default**. The assistant decides what to remember, phrases it in its own words, and uses it silently. You find out what it believes about you when an answer comes back subtly wrong. Every other failure below is worse because of this one: you cannot correct, move, or trust what you cannot read. **Memory you cannot see is memory you cannot manage.** ### Stale and Wrong Memories People change their minds; memories do not expire. A preference from six months ago, a project you abandoned, a guess the model made from one offhand remark: all of it can keep **quietly steering answers** long after it stopped being true. Worse, a wrong inference looks exactly like a right one. An AI that confidently remembers the wrong thing is less useful than one that asks. ### Locked to One Assistant Your ChatGPT memory does not follow you into Claude, and your Claude memory does not follow you into Gemini or your coding agent. Import tools help once, in one direction, and only into the vendor offering them. Most people now use **several assistants** for different jobs. With memory siloed per vendor, every switch starts from zero, and the context you built up becomes a reason not to use the best tool for the task. ### Memory Without Sources A saved memory says *"prefers Postgres"* but not **where that came from**: a benchmark you ran, a colleague's opinion, a throwaway line. Once a claim loses its source it cannot be checked, weighed or retracted. For personal preferences that is tolerable. For research, decisions or anything you might defend later, **a memory with no source is a rumour** your assistant repeats with confidence. ### It Remembers You, Not Your Thinking Built-in memory is optimized for **personalization**: facts about you that make the next answer fit better. It is not built to keep the **substance of your work**: the decision you reached in a long session, the evidence for it, the question you left open. That substance is exactly what gets lost in chat logs, and it is the part worth remembering. See [Your Best Ideas Are Dying in AI Chats](/m/ideas-dying-in-ai-chats). ## What Good AI Memory Needs ### Five Tests for Any AI Memory Before you trust a memory system with work that matters, run it through five tests. Each answers one of the failure modes on this map: 1. **Visible**: can you read everything it remembers? 2. **Editable**: can you correct and delete it? 3. **Portable**: does it work across every assistant you use? 4. **Sourced**: does each memory link back to where it came from? 5. **Structured**: does it keep how ideas relate, not just a list? Most built-in memory passes one or two. The test is not which product wins; it is **which memory you should keep yourself**. ### Visible: You Can Read It The first test is simple: **can you see all of it, in one place, in words you would recognize?** Not a settings list of fragments, but the actual picture the assistant is working from. Visibility is what turns memory from something done *to* you into something you own. It is also the only defence against the silent errors described in **The Memory You Can't See**. ### Editable: You Can Correct and Prune It Good memory is **gardened**. You should be able to fix a wrong fact, retire an old decision, and delete what no longer matters, as easily as you edit a document. And every change an AI makes to your memory should be **visible as a change**, so an assistant can never quietly rewrite what you know. ### Portable: Every Assistant Shares It Your memory should live **outside any one assistant**, so Claude, ChatGPT and your coding agent all read and write the same record. The **Model Context Protocol (MCP)** makes this practical: an MCP server exposes your memory as tools that any compatible assistant can call. Portable memory means switching assistants costs nothing, and what you worked out with one becomes context for the next. ### Sourced: Every Memory Links Back Each remembered claim should carry a **link to where it came from**: the conversation, the article, the document. That lets you check it, weigh it against newer evidence, and drop it when the source turns out to be wrong. Sources do not make a memory true. They make it **checkable**, which is the most you can ask of any record. ### Structured: Ideas Connected, Not Listed A flat list of memories cannot say that a decision **depends on** an assumption, or that new evidence **challenges** an old conclusion. Relationships are where reasoning lives. Structured memory, whether a knowledge graph or a visual map of connected thoughts, keeps the *why* next to the *what*, and lets you see at a glance which ideas hold the rest up. More in [Second Brain vs. Thought Networks](/m/second-brain-vs-thought-networks). ## Memory You Can See ### This Map Is a Memory You are reading an example of the alternative. This page is a **mindspace**: a visible map where every idea is a separate thought and every line states how two ideas relate. Nothing is hidden, everything is editable, and the structure is the point. Mindlify is built on that idea: **memory you can see, that you own, and that every assistant shares.** Not a replacement for built-in memory, which is fine for preferences, but a home for the thinking you actually want to keep. ### One Map, Every Assistant Mindlify is an **MCP server**, so Claude, ChatGPT, Claude Code and Gemini CLI can all read and write the same map. Decide something with Claude in the morning, and your coding agent can find it in the afternoon. Every AI write shows up on the canvas, and mindspaces are **private by default**. Setup for each assistant takes a few minutes at [mindlify.co/connect](/connect), and MCP access is included on the free plan. ### Ask Your Map Once your memory is a map, you can **ask it questions in plain language** and get answers grounded only in the thoughts it holds, with each source thought lit up on the canvas. When the map does not cover the question, it **says so** and offers to keep the gap as an open question, instead of filling it with a plausible guess. Memory that admits what it does not know is memory you can trust. ### Audit Your AI Memory This Week A practical first step, whatever tool you use: 1. Open the memory settings in **each assistant** you use and read what it has saved. 2. **Delete** anything stale, wrong, or that you would not want repeated. 3. Notice what is **missing**: the decisions and findings from your best sessions are rarely there. 4. Move those into one place **you** control, with a link to where each came from. Most people are surprised by step one, and more surprised by step three. ### From 100 Chats to One Memory The fastest way to build visible memory is to **start from the conversations you already had**. Paste a long chat or an article into Mindlify and it is unpacked into connected thoughts, each linked back to its source, and attached to what your map already holds. The full walkthrough is in [From 100 AI Conversations to One Visual Map](/m/from-100-chats-to-one-map). Mindlify is free to start: no credit card, and every account begins with seven days of Premium. ## How These Concepts Connect - **AI Memory, Explained** → **The Model Remembers Nothing** - **AI Memory, Explained** → **AI Memory Became a Category** - **AI Memory, Explained** → **The Memory You Can't See** - **AI Memory, Explained** → **Five Tests for Any AI Memory** - **AI Memory, Explained** → **This Map Is a Memory** - **The Model Remembers Nothing** → **Layer 1: The Context Window** - **The Model Remembers Nothing** → **Layer 2: Saved Memories** - **The Model Remembers Nothing** → **Layer 3: Retrieval (RAG)** - **The Model Remembers Nothing** → **Layer 4: Structured Memory** - **AI Memory Became a Category** → **ChatGPT Memory** - **AI Memory Became a Category** → **Claude Memory** - **AI Memory Became a Category** → **Gemini Memory** - **AI Memory Became a Category** → **Memory Layers for Agent Builders** - **The Memory You Can't See** → **Stale and Wrong Memories** - **The Memory You Can't See** → **Locked to One Assistant** - **The Memory You Can't See** → **Memory Without Sources** - **The Memory You Can't See** → **It Remembers You, Not Your Thinking** - **Five Tests for Any AI Memory** → **Visible: You Can Read It** - **Five Tests for Any AI Memory** → **Editable: You Can Correct and Prune It** - **Five Tests for Any AI Memory** → **Portable: Every Assistant Shares It** - **Five Tests for Any AI Memory** → **Sourced: Every Memory Links Back** - **Five Tests for Any AI Memory** → **Structured: Ideas Connected, Not Listed** - **This Map Is a Memory** → **One Map, Every Assistant** - **This Map Is a Memory** → **Ask Your Map** - **This Map Is a Memory** → **Audit Your AI Memory This Week** - **This Map Is a Memory** → **From 100 Chats to One Memory** ## Frequently Asked Questions ### What is AI memory? AI memory is any mechanism that lets an AI assistant carry information from one conversation into the next. Language models are stateless, so memory is built around them in layers: the context window (one conversation), saved memories (short facts about you), retrieval over past chats and documents (RAG), and structured memory that stores how ideas relate, such as a knowledge graph or a visual map of connected thoughts. ### How is AI memory different from the context window? The context window is working memory: everything in the current conversation, up to a token limit, and it is gone when the chat ends. AI memory is what persists after that, such as saved facts, searchable chat history or a structured knowledge store that the assistant can bring back into a future context window. ### Do ChatGPT, Claude and Gemini share memory? No. Each assistant keeps its own memory, stored with that vendor. Claude and Gemini offer tools to import context from other assistants, but memory does not stay in sync across them. To share one memory across assistants, keep it outside all of them, for example in an MCP server that Claude, ChatGPT, Claude Code and Gemini CLI can all read and write. ### What is an AI memory layer? An AI memory layer is infrastructure developers add underneath an AI agent so it remembers its users across sessions. Examples include Mem0, Zep (built on the Graphiti temporal knowledge graph), Letta (from the MemGPT research), Supermemory and Cognee. They solve how an app remembers its users, rather than how a person keeps and reviews their own thinking. ### Can I see and edit what my AI remembers about me? Partly. ChatGPT, Claude and Gemini each let you review and delete saved memories in their settings, with Claude putting the most emphasis on reading and editing. What an assistant infers from your chat history is usually not shown as an editable list. For memory you can fully read, correct and share across assistants, use a visible store you control, such as a Mindlify mindspace. ### What makes AI memory trustworthy? Five properties: it is visible (you can read all of it), editable (you can correct and prune it), portable (every assistant you use can access it), sourced (each memory links back to where it came from), and structured (it keeps how ideas relate, not just a list of facts). --- _Source: https://mindlify.co/m/ai-memory-explained. Published by [Mindlify](https://mindlify.co), AI-powered thought networks._