Secure External Memory Integration for ChatGPT – Persistent Context Across Conversations and Devices

ChatGPT Feature requests
1

Feature Request: Secure External Memory Integration and Persistent Context for ChatGPT

Hello OpenAI team,

I’m a ChatGPT Plus user and Android beta tester. I would like to propose a feature that could significantly improve long-term AI assistance: secure, user-controlled external memory integration.

The problem

ChatGPT can remember useful information, but long-running projects and personal assistant workflows often require more context than its current memory and conversation windows can reliably provide.

Important technical decisions, previous experiments, project milestones, and personal preferences can become difficult to recover across conversations.

Chat history may remain available when changing devices, but having access to old conversations is not the same as maintaining reliable, structured context.

Proposed solution

Please consider introducing an optional external memory system with the following capabilities:

  1. External storage connections: Allow users to connect supported storage providers or private databases, such as Google Drive or PostgreSQL, through secure, permission-controlled integrations.

  2. Persistent project memory: Allow ChatGPT to retrieve relevant facts, technical decisions, project progress, and previous results across conversations.

  3. Cross-device continuity: Make relevant context available when users switch phones, computers, or reinstall the application, using their authenticated account.

  4. User control: Let users review, correct, export, and delete stored information and decide which projects can access which memories.

  5. Security and privacy: Support encryption, explicit consent, limited permissions, access logs, and the ability to disconnect external storage.

  6. Reliable memory updates: Distinguish confirmed facts from hypotheses, plans, outdated information, and cancelled decisions to prevent contradictory memories.

  7. Automatic project summaries: Optionally generate structured progress summaries so complex work can continue without repeatedly reconstructing previous discussions.

Real-world use case

I’m working on a long-term personal AI assistant project involving Android integration, persistent memory, and technical testing.

A substantial amount of time is spent preserving context, transferring project information, and checking which decisions or tests were previously completed.

An official external memory interface could reduce repeated work and make ChatGPT much more effective as a long-term assistant.

Why this matters

For users who treat ChatGPT as a long-term collaborator rather than a one-time question-answering tool, continuity is essential.

The goal is not simply to store more chat history. It is to enable reliable, secure, portable, and user-controlled memory.

Thank you for considering this proposal and for continuing to improve ChatGPT.

— A ChatGPT Plus user and Android beta tester from Lithuania

2

Labels like confirmed, hypothesis and plan won’t tell retrieval which value is current when two entries about the same thing disagree, and an old “confirmed” entry can easily outrank a newer one and get surfaced as the truth.

Each entry also needs a timestamp and a source, plus an explicit link when a newer fact replaces an older one, so the old entry gets marked as superseded instead of just scoring lower. You already want users to review and correct memories, so conflicts between entries could show up in that same view for them to settle.

3

I already work with that everyday. Everything i do with GPT also generates or updates the relevant/necessary documents that are then re-used in the next chats/conversations.

It would be virtually impossible to work without these, relying only on the standard system context.

4

Timestamps and sources alone won’t settle it if retrieval just aggregates scores, so the supersession link matters more than the label. I’d add one thing: supersession should be directional and auditable, so if a user later decides the newer entry was wrong, the chain can be reversed without losing the original. Surfacing unresolved conflicts in the review view is a good call, but it needs a default when the user never responds, otherwise retrieval picks arbitrarily. What’s your view on whether superseded entries stay retrievable for history, or should they drop out entirely?

5

Thanks, Nicholas — I agree that supersession should be directional, reversible, and auditable.

My preference is to retain superseded entries in a versioned history, but exclude them from normal retrieval as current facts. They should remain accessible when the user explicitly asks about history, changes over time, or why a particular answer was given.

A practical design could include:

  • An explicit supersedes relationship linking the new entry to the previous one, with timestamps, provenance, and a reason for the change.
  • A clear state for each entry (current, superseded, disputed, or retracted), rather than relying on similarity scores or labels alone.
  • Retrieval that defaults to the latest verified, non-disputed entry, while flagging unresolved contradictions instead of silently choosing whichever scores highest.
  • A user-facing review trail that allows a mistaken update to be reversed without deleting the original evidence.

If the user never resolves a conflict, I would avoid treating either conflicting entry as certain. For low-stakes contexts, the assistant can acknowledge uncertainty; for consequential decisions, it should ask for clarification rather than invent a definitive current fact.

We’ve been exploring similar fact-versus-plan distinctions, versioning, and cancellation handling in a small independent memory prototype, so this is a particularly useful design discussion. Thanks for raising it!