Which AI Keeps Context Across Sessions? What ChatGPT, Claude, Gemini and Ocai Each Keep

★
By Mahesh ● Practical Guide ● tool comparison ● 16 min read ● 18 November 2026

Every competitor fact below was checked against provider documentation on 18 November 2026. Anything that moves fast carries a date next to it.

Quick Answer

Four products keep some of your context across sessions today. What separates them is which part of your context survives.

  • ChatGPT keeps saved memories and, when you switch it on, references your past chats. Projects add a scope where files and instructions travel together.
  • Claude keeps memory as a set of individual topics, searches your past chats on paid plans, and gives every project its own memory.
  • Gemini keeps memories from your past chats and standing instructions, on a Google AI plan, in the US, with each item expiring after 18 months at most.
  • Ocai keeps the state of the work itself: a project’s decisions, files, preferences and open threads.

If your problem is typing your role and your tone into every new chat, ChatGPT or Claude solves it. If your problem is re-briefing a project every Monday, that is a different layer, and it is the one Ocai is built around.

1. Four different jobs hide inside “context across sessions”

Say you spent three weeks on one piece of work. Nine files, a decision you reversed halfway through, and an agreement about how the writing should sound. You open a new chat. None of it is there. So you go looking for the AI that keeps context across sessions. The reason the answers online sound vague is that the question is four questions stacked on top of each other. Pull them apart and each one has a real answer.

The four layers, in plain terms

Layer one is what you are like. Your job, your tone, the format you always want. Preferences.

Layer two is what you already said. Last month’s thread, the pricing argument, the draft you threw out. Chat history, searchable.

Layer three is what this work is. The brief, the rules, the reference files, the conventions of one project. Project context.

Layer four is where the work stands. Which file changed, what got decided, what is still open. Working state.

A product can be strong on one layer and flat on another. That is why an assistant can remember that you like short answers and still have no idea you dropped Postgres on Thursday.

The mental load of re-explaining yourself is where that shows up first.

Layer What it holds What it costs you when it is missing Where it usually lives
1. Preferences Role, tone, format, dislikes Retyping the same three sentences Memory settings
2. Past chats What was said in earlier threads Losing the thread of an old argument Chat search or reference history
3. Project context Brief, files, standing instructions Re-uploading the same documents Projects or workspaces
4. Working state Decisions, reversals, open threads Rebuilding the job before you can start it Journal plus files

Layer four is where the hours go. If you have felt your assistant drift inside one long session, that is the same failure at a smaller scale. The anatomy of context window amnesia walks through the mechanism.

2. ChatGPT keeps preferences, and project context when you set it up

OpenAI’s help centre describes the job plainly. When memory is on, ChatGPT “can remember relevant preferences and details from your chats and other available sources”. The stated benefit is that you “spend less time repeating yourself”.

The sources list is wider than most people assume. Depending on plan and region, it can include past chats, saved memories, custom instructions, saved files, and content from connected apps such as Gmail.

Two controls decide what ChatGPT keeps

Saved memories are the explicit items. You ask it to remember something, or it saves a detail on its own. They are stored separately from chat history, and OpenAI is clear that deleting the original chat does not delete the memory that came from it.

Reference chat history is the broader control. Turn it on and ChatGPT can use relevant material from past conversations.

Turn it off and OpenAI says the information remembered from past chats “is scheduled for deletion from OpenAI systems within 30 days”. Your chats themselves stay where they are.

There is also a newer split inside ChatGPT. Improved memory keeps a rolling summary of relevant context, and legacy saved memories keep a list of individual items.

You can switch between the two experiences in settings.

The projects feature is described as keeping “related chats, files, and instructions together”. Project instructions override your global ones inside that project. There is a useful detail for anyone who shares work: shared projects use project-only memory, so they cannot read a member’s personal memories.

The limit is stated by OpenAI as well. Memory “does not retain every detail from every conversation”, and what it holds can change as your context changes. Availability varies by plan, region, platform and workspace settings.

For the field-by-field version, including models and pricing, the full OpenCraft AI vs ChatGPT comparison is the companion read.

3. Claude keeps topics, searches your chats, and scopes memory to each project

Anthropic’s documentation is the most specific of the three about what memory is made of. Claude “saves memory as a set of individual topics as you chat, rather than summarizing conversations after they end”. Mention that a deadline moved, and the next conversation already knows.

Chat search is a separate feature with a separate rule. Ask Claude to look through earlier conversations, and those searches “use Retrieval-Augmented Generation (RAG) and will appear as tool calls”.

Search reaches all chats outside projects, plus the conversations inside whichever project you are in. It does not cross between projects.

Retrieval over your own chat history is a different job from memory, and the two fail in different ways.

Memory itself is on by default on the Free, Pro, Max and Ultra plans, on web, desktop and mobile. On Team and Enterprise, the owner decides. In most organisations created on or after 31 October 2026, memory starts on for each member unless they turned it off before.

Pause keeps memory, reset deletes it

  • Pause memory keeps what Claude already knows. It stops the use of memory and stops new entries, and conversations during the pause are not added later.
  • Reset memory deletes everything, including project memories, and cannot be undone.

Each project carries its own memory. Claude uses that project memory alongside your account memory, but not memory from your other projects. If you want a project sealed off completely, turn off “Use account memory” on the project page.

There is one more rule worth knowing. Claude does not store sensitive topics by default. Some information is never saved even when asked, including government ID numbers, financial account numbers and immigration status.

Memory also travels between chat and Claude Cowork when Cowork runs in the cloud. That is the closest Anthropic comes to one memory across surfaces.

4. Gemini keeps memories from past chats, inside hard limits

Google’s help page states the purpose in one line. Gemini can remember “specific facts, interests, and preferences across conversations so you don’t have to repeat info in new chats”. It “will remember the information or instructions you share with it, even in new chats”.

Two things sit under that promise, and the names are easy to mix up.

Memory is the feature that carries things across chats. It is the switch you turn on or off under Settings, then Personal Intelligence.

Instructions for Gemini are standing rules you set once at gemini.google.com, under Settings and then Personal context. Google’s privacy pages call these “Saved Info” in some locales. They apply to every chat.

The limits are unusually well documented, so here they are without softening. You need to be over 18, on a Google AI plan, signed in with a personal Google Account.

Work, school and supervised accounts do not get the feature. It is available in the US only for now.

The 18-month limit

Then there is the clock. Every item you ask Gemini to remember, plus every instruction you give it, “will last for a maximum of 18 months”.

You can correct a wrong memory in plain language and it updates. After 18 months it is gone, and you add the detail again.

5. Ocai keeps the working state of the project

The first four layers are the ones the big assistants compete on. Ocai is aimed at the fourth.

Its product page puts it this way. It “remembers your work across sessions, so the next prompt starts sharper”, and it learns “how you work, files, context, decisions”.

Under that, the mechanism is a project. A named project carries its own preferences, memory, files and conventions.

What a named project actually holds

Memory sits at project level, with three parts doing different jobs. A journal records what happened and what was decided, and semantic recall finds the thread of an earlier conversation without you naming it.

Preferences layer globally and per project, with projects isolated so switching work does not leak context.

Files are what separate layer three from layer four. Ocai stores work in collections, and every save is a git commit, so a later session can look at what changed instead of being told. Code runs in an isolated sandbox, so the file the assistant wrote last week still exists this week with its history attached.

That is why the pricing model is credits rather than a seat. OpenCraft publishes plans and per-model rates: 0.5 credits on Ocai Swift up to 26 on Kimi K3, per million input tokens. There are no rate limits, and credits do not expire.

One caveat about how this works. Project memory helps when the work is organised into a project. Opened as a loose chat, it behaves like any other assistant.

6. Side by side: which context survives a new chat

Each column has a field-by-field page: ChatGPT and Claude.

Question ChatGPT Claude Gemini Ocai
Remembers preferences across chats Yes, memory plus custom instructions Yes, as memory topics Yes, memory plus instructions Yes, global and per-project
Can search earlier conversations Yes, with reference chat history on Yes, on paid plans, inside project boundaries Carries memories forward, not described as search Yes, semantic recall
Project-scoped context Projects hold chats, files and instructions Project memory, separate per project Gems and instructions Projects with memory, files and conventions
Working state of the task Not described as a stored state Not described as a stored state Not described as a stored state Journal plus git-versioned files
Documented retention limit Chat-derived memory deleted within 30 days of switching it off Project memory deleted with the project Items expire after 18 months at most Credits never expire, file history is versioned
Where you manage it Settings, Personalization, Memory Settings, Your Claude, Manage Settings, Personal context Project settings

Read the table by row. Almost every product now wins row one. Fewer win row four.

7. When the context gets long, something always gets dropped

No assistant ships an infinite window, and several say so themselves.

The vendors are explicit about it. OpenAI states that memory does not keep every detail, and Anthropic describes memory as topics chosen per chat.

Google goes furthest and puts a date on it. Saved memories expire after 18 months at most.

So the useful question changes shape. Does it keep the frame when it has to cut, and can you see the cut?

That question has a design answer. The difference between retrieval and agent memory is where the two systems stop being the same thing. Retrieval reads a corpus you curated. Memory decides what deserves to exist in the first place, and that deciding happens on every turn.

8. The ten-minute test for any AI that claims cross-session memory

You do not need a trial period to answer the buyer question. You need five questions and a stopwatch. Run them in this order, on whichever assistant you are considering.

Test 1, the cold open. Start a new chat with zero preamble. Ask what you are working on and what the next step is. If you spend more than a minute explaining, layer four is missing.

Test 2, the reversal. Change a decision on purpose, then open a new chat and ask which version is current. A system that keeps both versions and hands you the old one has no rule for replacing a memory. That failure is worse than silence, because it answers with confidence.

Test 3, the artefact. Ask which file it produced last time and what changed in it. Preferences and chat search cannot answer this. Only stored files with history can.

Test 4, the clock. Ask what it remembers about you and when that expires. If the answer is “forever” and no documentation backs it, treat the claim as marketing until you find the page.

Test 5, the audit. Can you list what is held, edit it, and delete one item without deleting its source? Can you delete the source without leaving the item behind? Both directions matter, and the second one is where products get quiet.

Score each test zero or one. Four or five, and the assistant holds context the way you need it. Below three, you will be the memory subsystem, and you will be doing it for free.

9. Which AI to pick, by what you want to stop repeating

Match the product to the layer you actually lose time on.

What you keep retyping Layer What to check before you buy
Your role, tone and format rules 1 Preference memory that is visible and editable
The thread of an earlier conversation 2 Chat search, and whether it is on your plan
A project’s files and standing rules 3 Project scope, file limits, isolation between projects
The state of the work from last week 4 A journal, versioned files, and recall that is not keyword based

The cases split three ways. If you mostly need an assistant that knows your name, your tone and your standards, ChatGPT and Claude both do it well. Claude’s per-project memory is the tidier of the two for agency work.

If you are inside Google’s ecosystem, Gemini fits the surrounding tools. Go in knowing the US-only and 18-month constraints.

If the thing you keep rebuilding is the job itself, that is layer four. The Claude comparison and the ChatGPT comparison show where Ocai sits against each on memory, models and pricing.

10. Where each one still falls short on context

None of these products is finished, and the gaps are documented by the vendors themselves.

ChatGPT’s memory availability changes with plan, region, platform and workspace settings. A feature you read about may not appear in your settings.

Claude’s chat search needs a paid plan, and searches stay inside project boundaries. That is right for privacy and awkward when a decision made in one project applies to another.

Gemini’s feature needs a Google AI plan, a personal account, and a US location as of 18 November 2026. Memory items last 18 months at most.

Ocai’s limits are equally specific. Project memory asks you to work in projects, and per-model context windows are not published as documented figures.

There is no public API, no mobile app, and integration is file-based rather than through SaaS connectors. If your work needs the assistant inside Drive or Slack, that gap is real.

Every one of those is a fact from a vendor page. The right product depends on which layer costs you the most.

Frequently Asked Questions

Which AI keeps context across sessions without me re-explaining?

All three major assistants now do, partially. ChatGPT keeps memories and can reference past chats, as the ChatGPT comparison records field by field. Claude keeps memory topics, searches past chats on paid plans, and keeps project memory separate.

Gemini keeps memories from past chats and standing instructions, on a Google AI plan in the US. Ocai keeps project-level working state, including a journal of decisions and versioned files.

Does a bigger context window solve this?

It delays the first cut. It does not cross the gap between sessions, because the window lives inside one request, which the anatomy of context window amnesia sets out step by step.

Closing a tab and coming back starts from whatever was stored outside the window. OpenAI says plainly that memory does not keep every detail, which is the same trade from the other side.

Does ChatGPT remember across all chats?

OpenAI’s wording is “across all chats” rather than across sessions, and saved memories are not tied to a specific conversation.

Two exceptions matter. Temporary chats do not create or update memories even with personalization on. Deleting a chat also does not delete a saved memory created from it.

Does Gemini memory expire?

Yes. Google states that anything you ask Gemini to remember, plus any instruction you give it, lasts a maximum of 18 months.

The same page lists the other conditions: over 18, a Google AI plan, a personal account, and US availability.

Can Claude keep a whole project in memory?

Yes, at project level. Each project has its own memory, and Claude uses it alongside your account memory but not memory from other projects.

Turn off “Use account memory” on the project page if you want that project fully separate. Memory is shared with Cowork only when Cowork runs in the cloud.

What happens to context if I turn memory off?

It depends which switch you use. In ChatGPT, turning off memory does not delete your past chats. Turning off reference chat history schedules chat-derived memories for deletion within 30 days.

In Claude, pausing keeps existing memory but stops new entries. Resetting deletes everything, including project memories, permanently.

Can I see and edit what an AI remembers about me?

ChatGPT shows a memory summary plus a sources list under responses, with controls to correct or delete. Claude lists what it remembers under Topics, and edits apply to every later conversation.

Gemini lets you correct a detail in chat or manage it under Personal context. If a product cannot show you its memory, it cannot be audited.

Is any of this private?

Each vendor publishes its own policy, and they differ. That is why the primary policy is the thing to read rather than a summary.

Two documented behaviours are worth knowing. Claude does not store sensitive topics by default and never stores some identifiers. ChatGPT lets workspace owners control memory for their organisation.

Sources


Start with free credits, no card needed, and run the five tests above on your own work before you decide anything.

Scroll to Top