Nightcap AI, digital life, and the tools changing how we connect

openai memory

OpenAI Memory Explained: ChatGPT Saved Memories, Dreaming, and Memory for Developers

The phrase openai memory covers two different things: the memory feature inside ChatGPT, and the memory you have to assemble yourself when building on OpenAI's models. This guide separates them, traces the documented timeline from the 2024 launch through the June 2026 dreaming release, maps which memory store applies on which surface, and ends with a checklist of where the sources agree and a decision tree for your situation.

By Nightcap Research Desk · Published September 1, 2026 · Last verified September 1, 2026

This article was researched with AI assistance and independently reviewed by multiple AI models before publication.

Key takeaways

  • OpenAI's June 4, 2026 post says memory first launched in April 2024 as saved memories, a feature that let you ask ChatGPT to remember information and carry it forward into future chats.
  • The same post describes dreaming as a background process introduced in April 2025 that curates and synthesizes ChatGPT's memory state from chat history, rather than waiting for explicit remember-this requests.
  • OpenAI's ChatGPT memory documentation states that ChatGPT web uses ChatGPT memory while local Codex clients use a separate local memory store and controls, and that the feature is managed from Settings > Personalization.
  • OpenAI's update notes say free users get a lightweight version providing short term continuity across conversations, while memory for Plus and Pro users provides a longer term understanding of the user.
  • On the developer side, an April 2024 OpenAI forum reply stated the API did not then offer a memory function, and another participant in that thread wrote that you pretty much have to build your own memory design.
  • A separate September 2024 OpenAI forum thread is where participants describe do-it-yourself patterns: an agent trained on Postgres, SOLR as a secondary option, and a tiered short/medium/long-term scheme with a consolidation step at a size threshold.

OpenAI Memory Explained: ChatGPT Saved Memories, Dreaming, and Memory for Developers

Searches for openai memory usually hide two different questions. One is about the product: how does the thing inside ChatGPT that remembers your preferences actually work, and how do you control it? The other is about engineering: how do you give an application built on OpenAI's models something that behaves like memory?

The published evidence for those two questions comes from different places — OpenAI announcement posts and product documentation for the first, developer forum threads and third-party tooling for the second. This guide keeps them separate, and flags where a statement is OpenAI's own, where it is a dated observation from a public forum, and where it is inference.

1. How ChatGPT's memory has evolved

Two OpenAI pages carry the timeline, and they were published years apart.

OpenAI's post Memory and new controls for ChatGPT is dated February 13, 2024 and carries a running stack of update notes on top of it. Those notes record that memory became available to ChatGPT Free, Plus, Team, and Enterprise users on September 5, 2024; that memory in ChatGPT became more comprehensive on April 10, 2025; and that as of a June 3, 2025 update, memory improvements were starting to roll out for free users, with ChatGPT referencing your recent conversations to provide more personalized responses in addition to the saved memories that were there before.

OpenAI's later post, Dreaming: Better memory for a more helpful ChatGPT, is dated June 4, 2026 and is framed as improving memory synthesis in ChatGPT to optimize for freshness, continuity and relevance. It gives the origin story directly: memory first launched in April 2024, also known as saved memories.

The two pages stamp different dates on nearby events — one is an announcement page accumulating update notes, the other a later release post looking back. Inference: read together they describe one continuous product line rather than competing histories, but if you need an exact date for a specific capability, take it from the page that announces that capability rather than reconciling the two.

2. The two mechanisms: saved memories and dreaming

OpenAI's dreaming post distinguishes the two clearly.

Saved memories are the original mechanism. The feature let you ask ChatGPT to remember information and carry it forward into future chats — an explicit, user-initiated write.

Dreaming is the automatic one. According to OpenAI's description, in April 2025 OpenAI updated ChatGPT's memory by giving the model the ability to reference chat context outside of the saved memories list, introducing the first version of dreaming — a method for ChatGPT to automatically curate memories in the background by referencing chat history. In contrast to saved memories, dreaming leverages a background process that allows ChatGPT to learn from many conversations and synthesize ChatGPT's memory state in order to provide the freshest, most relevant context to your conversations. OpenAI adds that dreaming also makes it easier for memory to include context that occurs naturally in conversation, without relying on explicit requests to remember something.

That second point is the practical difference for a user: with saved memories you decide what gets stored; with dreaming, ordinary conversation is a source.

3. Memory is not one store — it depends which surface you're on

This is the part most overviews skip, and it matters if you use more than one OpenAI client. OpenAI's memories documentation draws the boundaries:

  • ChatGPT web uses ChatGPT memory, while local Codex clients use a separate local memory store and controls.
  • ChatGPT Work uses the memory settings available to your account and workspace; it does not use a local Codex memory store or local memory controls.
  • The IDE extension uses the connected Codex host's local memory store.
  • Computer History is described as a macOS desktop feature that turns activity across allowed apps and websites into memories and a timeline that ChatGPT and Codex can reference.

The same documentation explains that once you enable memories, Codex can turn useful context from eligible prior chats into local memory files. So turning memory off is not necessarily a single global switch across every OpenAI surface you touch — the docs describe distinct stores with distinct controls.

4. Controls, and one dated criticism

On control, the documentation is short and specific: manage the feature from Settings > Personalization when you need to turn it on or off, per OpenAI's memories docs. OpenAI's help center also hosts a Memory FAQ page.

A public counterpoint is worth knowing about, with its date attached. In a March 2025 thread on OpenAI's developer forum about long-term AI companionship, a poster listed "No User-Controlled Memory Management" as a complaint, writing that unlike Gemini, which allows users to view, edit, or delete stored memory, ChatGPT operates in an opaque, automatic manner.

Two cautions on that. It is one user's characterization on a community forum, not an OpenAI statement. And it addresses a different level than the documented Settings toggle: viewing, editing and deleting individual stored items is a finer-grained question than switching the feature on or off. Treat it as a signal about what users wanted at that time, not as a current specification.

5. What differs by plan

OpenAI's update notes on the memory announcement page draw a tier distinction: free users have a lightweight version of memory improvements that provides short term continuity across conversations, while memory for Plus and Pro users provides a longer term understanding of the user.

The ChatGPT pricing page lists longer memory among the items in the Go plan, described as best for longer conversations and including more messages with tools, more uploads, more image creation and more voice chats on top of Free — with a note that this plan may include ads. Plan contents change; verify on the pricing page before deciding.

6. Memory for developers building on OpenAI

Here the picture is different, and the strongest public evidence is dated.

In an April 2024 OpenAI forum thread titled How do I enable or disable memory in API?, a developer asked whether the memory option could be toggled through the API. A replying community member stated that the API did not at that time offer a memory function, and another responded that you pretty much have to build your own memory design. That thread contains the build-it-yourself answer; the concrete implementation patterns cited below come from the separate September 2024 thread.

In Will Memory capabilities come to the API?, started September 2024, one participant described training an agent on Postgres and pointed to truncation strategy in the OpenAI docs; another offered SOLR as a secondary option; one developer running a sim-racing assistant described a tiered scheme where, once a message list crosses a size threshold, a memory consolidation function builds a payload combining short-term, medium-term and existing long-term memory. A participant in that thread also argued memories should always be available rather than spending extra internal calls on deciding whether what is in memory will be useful.

Those are individual developers writing in 2024 and early 2025, not OpenAI API specifications, and the API surface has moved since — the current OpenAI API documentation directs developers to start with the Agents SDK and to choose between GPT-5.6 Sol for complex reasoning and coding, GPT-5.6 Terra to balance intelligence and cost, or GPT-5.6 Luna for cost-sensitive, high-volume workloads. Check the current docs before assuming a 2024 forum answer still describes today's options.

Third-party memory layers also target this gap. A March 5, 2026 post from the Hindsight team, Give Your OpenAI App a Memory in 5 Minutes, describes building a ChatGPT-style chatbot with persistent memory using the OpenAI SDK plus their tool, via three API calls — retain(), recall() and reflect() — which they say lets an app remember users across restarts with no vector database or RAG pipeline required. Their described loop is: user message, recall(query) to pull relevant memories, an OpenAI completion with memory injected into the system prompt, then retain(exchange) to store the conversation. They install it with pip install hindsight-all and run a local server at http://localhost:8888. The motivation they give from building their own internal AI project manager is familiar: context windows fill up, token costs explode, you start truncating history, and the assistant forgets early decisions — so what you need is to store facts as they happen, retrieve only what's relevant, and synthesize when necessary. That framing of the problem is useful independent of whether you use their product; the capability claims are the vendor's own.

7. Checklist: what the sources agree on

  • Two write paths exist in ChatGPT. Explicit saved memories and automatic background synthesis are described as distinct mechanisms in OpenAI's dreaming post.
  • There is a documented on/off control. Settings > Personalization, per the memories documentation.
  • Memory availability is tiered. Free is lightweight short-term continuity; Plus and Pro get longer-term understanding, per OpenAI's update notes.
  • Stores are per-surface. ChatGPT web, local Codex clients, ChatGPT Work and the IDE extension are documented as using different memory stores.
  • Developer memory is assembled, not inherited. The two 2024 forum threads cited here point developers toward building their own design rather than flipping an API switch — the April 2024 thread for the reply that the API offered no memory function and that you build your own, the September 2024 thread for the concrete patterns.
  • Where to be careful: the two OpenAI pages carry different date stamps for adjacent events, and the March 2025 forum criticism about opaque memory management speaks to per-item view/edit/delete granularity rather than the documented on/off toggle — different questions, both worth checking against current docs.

8. Decision tree: find your situation

You want ChatGPT to stop remembering things. Go to Settings > Personalization, which the memories docs name as the place to turn the feature on or off — then check the other surfaces you use, since local Codex clients are documented as having a separate store and separate controls.

You want to know what changed recently. Read the dreaming release post for the background-synthesis model, then the update notes stacked on the original memory announcement for the rollout sequence.

You're on a free plan and want more continuity. OpenAI's notes describe free memory as lightweight short-term continuity versus longer-term understanding for Plus and Pro, and the pricing page lists longer memory under Go — compare those descriptions against what you actually need before paying.

You're building an app on OpenAI models. Start from the current API docs rather than older forum answers. If you do need to build it, the patterns publicly described in the September 2024 forum thread are a database such as Postgres, a search index such as SOLR, and a tiered short/medium/long-term scheme with a consolidation step at a size threshold; a dedicated memory layer such as the one in Hindsight's walkthrough is the vendor-supplied alternative.

You care about what's stored about you specifically. Start with the account-level controls in the memories documentation; OpenAI's help center also has a Memory FAQ page.

Sources