--- title: "How ChatGPT Memory Works" description: "How ChatGPT memory works as of September 2026: saved memories, reference chat history, the memory summary, what memory full means, and how to clear it." canonical: https://past.dev/blog/chatgpt-memory date: 2026-09-02 category: Research authors: The past.dev team --- # How ChatGPT Memory Works ChatGPT memory is the account-level layer that carries context between conversations. As of September 2026 it works like this: with memory on, ChatGPT automatically remembers useful context from your chats, uploaded files, and connected apps, and it maintains an editable [memory summary](https://help.openai.com/en/articles/8590148-memory-faq) of what it knows about you. This system began rolling out on June 4, 2026 and replaced the earlier pair of mechanisms, discrete saved memories plus a reference chat history setting, which survives as a legacy option ([release notes](https://help.openai.com/en/articles/6825453-chatgpt-release-notes)). Every claim below links to OpenAI's documentation. ## What ChatGPT memory stores and where to see it - Memory controls live in Settings > Personalization > Memory, where you can enable or disable memory at any time; when enabled, it draws on your chats, files, and connected apps ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). - The memory summary shows what ChatGPT remembers, updates automatically as you chat, and shows a last-updated time ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). It is a synthesis rather than a complete inventory: some details never appear in it, and the documented way to check a specific fact is to ask in chat ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). - To edit it, type the change into the text box at the bottom of the summary, or highlight any text to make a specific correction ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). - The book icon under a response lists the sources that personalized it (custom instructions, past chats, files, memories), each with an explanation of why it was used ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). - Memory is stored separately from chat history, so deleting a chat does not delete memories created from it ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). ## Saved memories and reference chat history The system most searchers ask about ran from 2024 to mid-2026 and remains available under Settings > Memory > Saved memories ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). OpenAI defined its two parts on April 10, 2025: "saved memories" you ask ChatGPT to keep, and "chat history", insights ChatGPT gathers from past chats to improve future ones ([announcement](https://openai.com/index/memory-and-new-controls-for-chatgpt/)). Each part has its own toggle, and OpenAI's projects documentation still names both: "Reference saved memories" and "Reference chat history" ([Projects in ChatGPT](https://help.openai.com/en/articles/10169521-using-projects-in-chatgpt)). The dated evolution: - September 5, 2024: memory reached Free, Plus, Team, and Enterprise accounts ([announcement](https://openai.com/index/memory-and-new-controls-for-chatgpt/)). - June 3, 2025: Free accounts got a lightweight version that references recent conversations for short-term continuity; in the EEA, Switzerland, Norway, Iceland, and Liechtenstein it is opt-in at Settings > Personalization > Memory > Reference chat history ([release notes](https://help.openai.com/en/articles/6825453-chatgpt-release-notes)). - October 15, 2025: ChatGPT began managing saved memories automatically, keeping relevant details prioritized and moving less important ones to the background ([release notes](https://help.openai.com/en/articles/6825453-chatgpt-release-notes)). - January 15, 2026: with reference chat history enabled, ChatGPT retrieves specific details from past chats more reliably, and any past chat it uses appears as a source ([release notes](https://help.openai.com/en/articles/6825453-chatgpt-release-notes)). - June 4, 2026: the current system began rolling out to reduce stale and contradictory saved memories, with increased memory capacity for Plus and Pro ([release notes](https://help.openai.com/en/articles/6825453-chatgpt-release-notes)). OpenAI's stated reason for the redesign is that manually managed memories went stale and contradicted each other, citing "I'm training for a marathon" coexisting with "I sprained my ankle" ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). Keeping [facts that change over time](/guides/facts-that-change-over-time) current is a general memory problem well beyond ChatGPT. Depth also varies by plan: the [pricing page](https://chatgpt.com/pricing) lists limited memory on Free, "Longer memory" on Go, "Expanded memory and context" on Plus, and "Maximum memory and context" on Pro. ## What ChatGPT memory full means and how to fix it The memory full notice belongs to the legacy saved memories system. That list has a fixed capacity: OpenAI's October 15, 2025 release note describes automatic memory management as a way to keep saved memories from "reaching capacity" and avoid the "memory full" state ([release notes](https://help.openai.com/en/articles/6825453-chatgpt-release-notes)). OpenAI does not publish the capacity itself, and users report that once the list is full, new memories stop being saved until space is freed ([OpenAI community thread](https://community.openai.com/t/chatgpt-plus-plan-memory-issue-no-warning-when-memory-is-full-leading-to-lost-entries/1111613)). Existing memories keep working either way, since saved memories are always considered in future responses until deleted ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). The documented fixes: - Delete individual saved memories you no longer need, or clear all of them, in settings ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). Deleting chats frees nothing here, because saved memories are stored separately from chat history ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). - Let automatic management work: since October 2025 it deprioritizes stale entries specifically to prevent the full state ([release notes](https://help.openai.com/en/articles/6825453-chatgpt-release-notes)). - Leave the legacy system: the June 2026 memory summary updates itself instead of filling a list, and Plus and Pro accounts received increased capacity ([release notes](https://help.openai.com/en/articles/6825453-chatgpt-release-notes)). Switch back and forth at Settings > Memory ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). ## Why ChatGPT forgets things inside one conversation Forgetting within a single chat is a different mechanism from cross-chat memory. Each model request runs inside a [context window](/glossary/context-window), a fixed token budget covering input, output, and reasoning tokens ([Understanding and counting tokens](https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them)). A conversation that outgrows the budget cannot pass all earlier turns to the next request: the window "limits the tokens a model can work with in a request" ([Understanding and counting tokens](https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them)). The [pricing page](https://chatgpt.com/pricing) documents the sizes as of September 2026: for default models ("GPT Instant") a total context window of 27K tokens on Free, 54K on Go and Plus, and 128K on Pro; for reasoning models ("GPT Reasoning") 256K on Go and Plus and 400K on Pro. Its footnote adds that ChatGPT manages a shared context window, and the space left for user input varies with the features in use and any memory content ([pricing page](https://chatgpt.com/pricing)). The two layers stay distinct: [cross-session memory](/glossary/cross-session-memory) persists between conversations, while the window is rebuilt on every request. ## Temporary chats, projects, and workspace controls - Temporary Chats start non-personalized: no memory, no custom instructions, and no new memories created; a personalized temporary chat can read existing memories but still creates none ([Temporary Chat FAQ](https://help.openai.com/en/articles/8914046-temporary-chat-faq)). Unsaved temporary chats stay out of history, though OpenAI may keep a copy up to 30 days for safety ([Temporary Chat FAQ](https://help.openai.com/en/articles/8914046-temporary-chat-faq)). - Projects offer default memory or project-only memory; with project-only, chats reference other chats in the same project and nothing crosses the boundary in either direction ([Projects in ChatGPT](https://help.openai.com/en/articles/10169521-using-projects-in-chatgpt)). Shared projects are always project-only ([Projects in ChatGPT](https://help.openai.com/en/articles/10169521-using-projects-in-chatgpt)). - On ChatGPT Business, memory is per account: workspace owners can disable it workspace-wide, and doing so deletes members' existing saved memories ([Memory FAQ, Business version](https://help.openai.com/en/articles/9295112-memory-faq-business-version)). Workspace data, including memories, is excluded from model training by default ([Memory FAQ, Business version](https://help.openai.com/en/articles/9295112-memory-faq-business-version)). - In ChatGPT for Healthcare and Enterprise regulated workspaces, improved memory is off by default and workspace admins choose who can enable it ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). - A log of deleted saved memories may persist up to 30 days for safety and debugging ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). ## What ChatGPT memory does not do Documented boundaries only: - It does not transfer between users: memories are tied to each account and cannot be shared with other members of a workspace ([Memory FAQ, Business version](https://help.openai.com/en/articles/9295112-memory-faq-business-version)). - Custom GPTs keep their own distinct memory, and builders never see yours ([announcement](https://openai.com/index/memory-and-new-controls-for-chatgpt/)). - Deleting a memory does not scrub past chats that mention it; those chats must be deleted separately ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). - Fully deleting a fact means removing every source it appears in: past chats, archived chats, files, the memory summary, and any connected apps that contain it ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). - The sources view may omit factors that shaped a response ([Memory FAQ](https://help.openai.com/en/articles/8590148-memory-faq)). ## For builders: this memory belongs to ChatGPT Everything above personalizes OpenAI's product for OpenAI's users; none of it is infrastructure your own assistant can call. If you are building an agent and need what ChatGPT's memory cannot give your product, past.dev is a memory API for that layer: POST /api/v1/ingest accepts raw text with its original timestamp (fields: content, timestamp), entity resolution links references to the same person across sources, and POST /api/v1/recall (field: query) returns ranked, dated evidence with sources and one of four status values (Supported, Conflicted, NoKnownSupport, UnknownBecauseDegraded). Facts carry event time, validity windows, and supersession links, and deployment is managed or self-hosted in your own Postgres. Start with the [Memory API quickstart](/docs/memory-api/quickstart) and see the [benchmarks](/benchmarks).