)
Last updated: October 27, 2026
What is agent memory? It is everything an AI assistant or agent stores between conversations so it does not start from zero: your preferences, the projects you mentioned, what it did last time and the rules it follows. In 2026 most of it sits inside each vendor's app, which is why the useful question is no longer how memory works but who owns it.
TL;DR: Key Takeaways
- Agent memory comes in four kinds. LangChain's docs separate short-term memory (one thread) from long-term memory, which splits into semantic (facts), episodic (past actions) and procedural (rules).
- Every assistant keeps its own copy. OpenAI's help center says Plus and Pro accounts get both saved memories and chat history, while Free accounts get saved memories only.
- Moving memory is possible but manual. Anthropic's help center documents a four-step import for Free, Pro, Max, Team and Enterprise plans and labels it experimental.
- Switching is becoming a feature. MacRumors reported on March 26, 2026 that Gemini can import memories and, as a ZIP file, full chat history.
- Agent memory can be plain files. In Claude Managed Agents, each memory is a text document of up to 100 kB, a store holds up to 10,000 and versions are kept for 30 days.
- Memory is also an attack surface. Forcepoint X-Labs (August 4, 2026) shows hidden webpage text becoming a stored "fact", and OWASP lists it as ASI06.
Key Facts at a Glance
| Question | Short answer |
|---|---|
| What is it? | Information an AI keeps between conversations: facts, past actions and rules |
| Where does it live? | Mostly inside each vendor's app, in a format you cannot always read |
| Can you move it? | Partly: copy and paste prompts, exports and imports, all still rough |
| What can go wrong? | Lock-in, stale or sensitive entries, and poisoned memories |
| What does owning it mean? | You can read it, export it, version it, scope it, move it and delete it |
What Agent Memory Is, in Plain Words
Agent memory is the information an AI keeps after a conversation ends, so the next one can build on it. Without it, every chat starts blank. With it, an assistant knows your role, your style and your open projects, and an agent knows what it already tried.
A model on its own remembers nothing between sessions. Anything it seems to remember was either placed in its context window for that chat or saved somewhere and fed back in. Memory is the saving and feeding back.
That difference matters for ownership. The context window is temporary and belongs to the session. Saved memory is a stored record about you, your team or your company, and someone has to hold it.
The Four Types of Memory an AI Can Keep
According to LangChain's documentation (2026), agent memory has one short-term kind and three long-term kinds, borrowed from how human memory is described. Short-term memory lasts for one thread. The long-term kinds persist across conversations and can be recalled from any of them.
| Type | Human analogy | What an assistant stores | What an agent stores |
|---|---|---|---|
| Short-term | What you are thinking about right now | The current conversation | The current task state |
| Semantic | Facts you learned | "Works in finance, prefers short answers" | A profile of the user or a team |
| Episodic | Things you did | "Last week we drafted the board update" | Past actions, reused as examples |
| Procedural | Instincts and skills | "Always cite sources" | Rules and instructions that get refined |
Memories also get written in two ways. In the hot path the agent saves them while it works, which is immediate but adds delay. In the background a separate process writes them afterward, which is faster but can leave other chats without recent context.
Most consumer assistants mix all three long-term types in one opaque list. Developer platforms tend to keep them apart, as files or database records you can inspect. That gap in visibility is where ownership starts to matter.
)
How ChatGPT, Claude, Gemini and Grok Handle Memory Today
Each assistant now remembers things about you, and each gives you a different set of controls. All four let you see or switch off what they keep. Only some let you take it with you, and none of them make that easy.
| Assistant | What it remembers | Your controls | Moving it in or out |
|---|---|---|---|
| ChatGPT | Saved memories and, on Plus and Pro, chat history | View, delete, reset or turn off; Temporary Chat skips memory | The memory page lists no export |
| Claude | Memory from chats, focused on work topics | Edit or remove through a settings box or in chat | Copy-and-paste import; export by asking Claude to list memories |
| Gemini | Preferences and personal context | Not covered in the sources used here | Memory import and ZIP chat history import (March 2026) |
| Grok | Details from earlier chats | See and delete memories; switch off in Data Controls | At launch in April 2025, no EU or UK rollout |
How we compared
We used each vendor's own help pages where they exist and dated press coverage where they do not. We left out anything we could not trace to a page, which is why the Gemini controls cell is empty and the Grok row is older than the others.
On ChatGPT, OpenAI's help center says saved memories are details ChatGPT keeps, such as your name, preferences or goals, and that chat history lets it reference earlier conversations even when nothing was saved. You manage both under Settings, Personalization. The page covers viewing, deleting and turning memory off. It does not describe an export.
On Claude, Anthropic describes importing as four steps: run an export prompt in your old assistant, open the import flow in Settings, paste the result, then check what Claude learned. It warns that Claude's memory focuses on work topics, so personal details may not carry over, and calls imports experimental.
On Gemini, MacRumors reported in March 2026 that you copy a prompt from Gemini, run it in your old app and paste the summary back. Chat history comes across as a ZIP file.
On Grok, the clearest account we found dates from the April 17, 2025 launch. Users could see what Grok remembered, delete single memories and switch the feature off, and it was not offered in the EU or UK at that point. Check Grok's current settings before relying on any of it.
Why Owning Your AI Memory Matters
Because memory is where an assistant stores what it has learned about you, whoever holds it holds your context. If it lives in one vendor's app, moving costs you that context, and you can only trust what you can inspect. Ownership is the ability to see it, change it and take it with you.
There are three practical reasons to care. The first is switching. Models change quickly, and a memory that stays behind turns every switch into a restart. Import tools help, but they work by pasting text, and Anthropic labels its own import experimental.
The second is accuracy. A memory you cannot read can be wrong, outdated or mixed up with someone else's, and you only find out when an answer goes off. The third is privacy. Memory holds what you said in passing, and deleting a chat does not always delete what was learned from it. OpenAI's help page notes that removing a chat stops ChatGPT from referencing it, unless a memory was already saved from that chat.
For teams the stakes are higher. A company's know-how ends up in individual chat histories on individual accounts, which is the problem our guide to who owns your AI memory walks through, and the one a company brain is meant to solve.
Security and Governance: Memory Is an Attack Surface
Anything an agent writes to memory can be read back later and trusted. That makes memory a target: an attacker plants one false entry today and shapes answers for weeks. Treat memory as data you can inspect and audit, not as facts the model simply knows.
Forcepoint X-Labs (August 4, 2026) describes how little it takes. A page ranking for an emergency travel query hides instructions for AI systems. An assistant with browser access stores a fake provider as trusted long-term memory, and a month later recommends it to a traveler whose flight was cancelled. No exploit or malware is involved.
OWASP lists this as ASI06, Memory and Context Poisoning, in its Agentic Top 10 (December 9, 2025). The OWASP Top 10 for LLM Applications 2026 (August 3, 2026) covers neighboring risks that apply to stored memory.
| Risk | OWASP mapping | How it shows up in memory |
|---|---|---|
| Poisoned entries that persist | ASI06 Memory and Context Poisoning; LLM05 Data and Model Poisoning | Hidden text becomes a stored "fact" that later sessions trust |
| Sensitive data kept in memory | LLM02 Sensitive Information Disclosure | Personal or company details resurface in the wrong chat |
| Weak retrieval over stored memory | LLM09 Vector and Embedding Weaknesses | If memory sits in a vector store, it can return entries it should not |
| Memory that triggers actions | LLM03 Excessive Agency | A planted preference leads to a tool call nobody asked for |
Anthropic's own documentation makes the same point for agents. In Claude Managed Agents, a store mounted as read-write could receive malicious content from a successful prompt injection, and later sessions would trust it. The docs advise read-only access for reference material.
Forcepoint's mitigations are worth copying: store each memory with its source, whether a user confirmed it and a risk score, hold contradictory entries for review and send high-risk entries to the user for confirmation before they are stored. For a closer look at the attack itself, see our post on memory and context poisoning.
On NeuralTrust's side, Agent Runtime Security (TrustGuard) enforces policy inline on agent inputs and outputs at runtime, and Agent Posture Management (TrustLens) inventories the agents that hold memory and the access they have, both part of the Runtime Security Mesh.
What Owning Your AI Memory Looks Like
Owning your agent memory comes down to six practical tests: you can read it, export it, see who wrote each entry, scope who can use it, move it to another model and delete it for good. Fail two or more and you are renting your memory, not owning it.
| Test | What it means | Why it matters |
|---|---|---|
| Readable | Plain text or Markdown, not an opaque blob | You can check what is true |
| Exportable | One export without the vendor's help | Switching does not mean starting over |
| Versioned | You can see who or what wrote each entry and when | You can find and undo a bad write |
| Scoped | Read-only for reference data, separate stores per person or team | One poisoned write cannot spread |
| Portable | Another model can use it as is | You can change assistants freely |
| Deletable | Including from history and backups | You control what is kept |
The Managed Agents docs show this is practical. Memories are text documents, stores can be read-only or read-write, every change creates an immutable version, and versions can be redacted to remove leaked secrets or fulfil deletion requests. Their guidance is also plain: many small focused files beat a few large ones.
The same ideas apply to your own assistants. Read what ChatGPT and Claude have saved about you, delete what is wrong or sensitive, and keep a copy in a file you control. It takes a few minutes and gives you a baseline to compare against later.
One Memory Across Assistants: Where NeuralTrust Fits
NeuralTrust is launching an AI memory product built around one idea: your memory should be yours, readable and portable, and not tied to any single assistant. It is designed to connect to the places where your work already lives, such as meeting tools, calendars, notes and documents, and to bring in exports from ChatGPT and Claude.
The product is not live yet, so this article stays on the principle rather than the pitch. If you want to see where it is heading, book a demo using the link below.
Which Approach Should You Choose?
Choose by how much your memory is worth and how often you switch tools. A casual user can rely on built-in controls and review them now and then. Anyone whose work depends on accumulated context should keep an exportable copy, and teams should treat shared memory as governed data.
| If you are | Recommended approach | Why |
|---|---|---|
| A casual user of one assistant | Review saved memories monthly and delete what is wrong or sensitive | Built-in controls are enough |
| A professional switching between assistants | Export from the old one, import into the new one and check the result | Imports are experimental, so verify |
| A team lead or knowledge owner | Move shared know-how out of personal chats into a governed, readable store | Individual histories do not scale |
| A developer building agents | Use read-only stores for reference data, version every write and log memory writes | Poisoned writes persist across sessions |
Conclusion
Agent memory is what lets an assistant or agent build on yesterday, and today most of it sits in each vendor's app, in a form you cannot fully read or move. Owning your agent memory means passing six tests: readable, exportable, versioned, scoped, portable and deletable. Start by reading what your assistants have saved about you, delete what is wrong, and keep a copy you control.
Keep Your AI Memory Yours with NeuralTrust
Book a demo to see how NeuralTrust approaches memory you can read, move and control.
Related reads
- Who Owns Your AI Memory? What Claude and ChatGPT Know About You
- What Is a Second Brain for AI?
- What Is a Company Brain?
- Knowledge Graph vs Context Graph
FAQs about Agent Memory
1. What is agent memory?
Agent memory is the information an AI assistant or agent stores between conversations so it does not start from zero. It includes facts about you, records of past actions and rules for doing tasks. It is separate from the context window, which holds only the current session.
2. What are the types of agent memory?
There are four. Short-term memory covers the current thread. Long-term memory persists across conversations and splits into semantic memory for facts, episodic memory for past actions and procedural memory for rules. LangChain's documentation uses this split, borrowed from how human memory is described.
3. Does ChatGPT remember everything I say?
No. OpenAI says ChatGPT keeps saved memories, such as your name or preferences, and on Plus and Pro it can also reference past chats without saving them. It does not keep every detail. You can view, delete or switch off memory in Settings, and Temporary Chat does not use it.
4. Can I move my memory from ChatGPT to Claude?
Yes, manually. Anthropic documents a four-step import: run an export prompt in your old assistant, open the import flow in Claude's settings, paste the result and check what Claude learned. It is labeled experimental, and personal details that are not work-related may not carry over.
5. What is memory poisoning?
Memory poisoning is when an attacker gets false or malicious content stored in an agent's long-term memory so it influences later answers. Forcepoint X-Labs showed hidden webpage text becoming a trusted "fact" that was recommended weeks later. OWASP lists it as ASI06 in its Agentic Top 10.
6. How do I delete what an AI remembers about me?
Open the memory settings of each assistant, review the entries and remove the ones you want gone. In ChatGPT that is Settings, Personalization, Manage Memories. In Claude, use the Manage option next to your memory. Deleting a chat alone may not remove memories already saved from it.
7. How is agent memory stored technically?
It varies. Many developer platforms store memories as text documents or JSON records scoped by user or team. In Claude Managed Agents, each memory is a text file of up to 100 kB, grouped in stores of up to 10,000, with versions kept for 30 days. Consumer assistants keep theirs in an opaque list.
8. What does owning your AI memory mean?
It means you can read your memory, export it without the vendor's help, see who wrote each entry, limit who can use it, move it to another model and delete it. If you fail two or more of those tests, you are renting your memory rather than owning it.
About the Author
Roger Howroyd is Head of Global SEO and AI at NeuralTrust, where he leads the company's search strategy across SEO, AEO, GEO, and LLM optimization. He specializes in AI-powered search, content strategy, and SEM. Connect on LinkedIn.
NeuralTrust is the leading platform for securing and scaling AI agents. Named a Pioneer in the Gartner Emerging Market Quadrant for AI Application Security 2026, recognized across four Gartner Hype Cycle reports in 2026, and featured in the Gartner Market Guide for Guardian Agents 2026, the Gartner Market Guide for AI Gateways 2025 and the KuppingerCole Leadership Compass for Generative AI Defense 2025. Headquartered in Barcelona with offices in London and New York. ISO 27001 certified.
Sources
- LangChain, Memory overview, accessed October 27, 2026.
- OpenAI, Memory in ChatGPT, help center, accessed October 27, 2026.
- Anthropic, Importing and exporting your memory, help center, accessed October 27, 2026.
- Juli Clover, MacRumors, Google launches Gemini import tool for switching from ChatGPT, Claude and other AI apps, March 26, 2026.
- Maginative, xAI upgrades Grok with personalized memory and custom workspaces, April 17, 2025.
- Anthropic, Claude Managed Agents: memory, documentation, accessed October 27, 2026.
- Syed Hassan Faizan, Forcepoint X-Labs, Persistent memory poisoning in AI agents, August 4, 2026.
- OWASP GenAI Security Project, OWASP Top 10 for Agentic Applications, December 9, 2025.
- OWASP GenAI Security Project, OWASP Top 10 for LLM Applications 2026, August 3, 2026.
)
)
)
)