For the first time, all three major consumer AI labs ship persistent memory — Anthropic with Claude Dreaming (announced May 6, 2026 at Code with Claude), OpenAI with ChatGPT Memory (now a year mature), and Google with Gemini Spark’s account-level context (launched May 19 at I/O 2026). The three implementations look similar on the marketing page. They are completely different products underneath.
If you’ve ever had ChatGPT keep recommending the wrong solution because it “remembered” a preference you actually wanted to change, or watched Claude forget the same project context for the eighth time this week, the mental model behind each system matters more than the marketing. This post is the cross-vendor comparison that nobody published yet because it required all three to ship — and only as of this week, they all have.
The three mental models
Claude Dreaming = offline curation. Memory updates happen as a background job that runs between your sessions, not during them. Anthropic’s framing: dreaming “reviews past sessions to find patterns and help agents self-improve.” The system reads up to 100 past session transcripts, identifies recurring workflows and mistakes, prunes duplicates, and writes the result to a separate memory store you can review before promoting. The model isn’t remembering you in real time — it’s distilling your past into a curated playbook.
ChatGPT Memory = explicit storage + implicit reference. OpenAI maintains two parallel mechanisms. Saved memories are structured facts ChatGPT stores about you (your name, your job, your preferences) that you can see and edit in Settings → Personalization → Memory. Chat history reference is the broader background context where ChatGPT pulls from prior conversations even when you didn’t explicitly save anything. Two systems, two control panels, two toggles. Most users don’t know there’s a difference.
Gemini Spark memory = unified account context. Google’s design is the broadest: rather than a memory layer that sits over conversations, Gemini Spark’s memory is built into your Google account itself — pulling from your Gmail, Calendar, Docs, Drive, search history, and prior Gemini chats as a unified context that any Gemini surface (mobile, web, Spark agent, Daily Brief) can read. There’s no separate memory feature to enable. Spark just knows what’s in your Google account because that’s where the memory lives.
These are not three flavors of the same thing. They’re three philosophies about where memory should live and who should curate it.
How Claude Dreaming actually works
Dreaming is a three-phase pipeline that runs asynchronously, between agent jobs:
Phase 1 — Read. The dream job takes inputs: an existing memory store (the agent’s current working memory), up to 100 past session transcripts, and an optional instructions field (max 4,096 characters) describing what to look for. The 100-session cap is a hard ceiling per dream operation.
Phase 2 — Curate. Claude analyzes the inputs for four categories:
- Workflows that the agent converges on across different jobs
- Error patterns — recurring mistakes plus successful recovery strategies
- Preferences — shared formatting, approval flows, tool choices across a team
- Coordination patterns — how multiple agents hand off in multi-agent setups
It also resolves contradictions (memory rot — stale or conflicting entries) and merges duplicates.
Phase 3 — Output. The curated result lands in a separate memory store, not the agent’s live working memory. You decide what happens next. In automatic mode, the new store replaces the old one without review. In review mode, a human approves or rejects individual memory edits before they affect the agent’s behavior. Most enterprise deployments choose review mode at first.
The Harvey result from Anthropic’s launch coverage — 6× task-completion rate improvement after enabling Dreaming on the legal AI workflow — is the most concrete number any of the three labs has published. The driver: the agent stopped forgetting filetype quirks and tool-specific workarounds between sessions. The pattern matters more than the number — Dreaming’s value is highest when agents repeat similar failures over time, lowest for one-off questions.
Access status: Research preview only as of May 21, 2026. You request access through an Anthropic form (no public signup). The API requires two beta headers: managed-agents-2026-04-01 and dreaming-2026-04-21. Claude Code users with version v2.1.59+ get the /dream command that triggers a dream run on their agent’s memory file.
How ChatGPT Memory actually works
OpenAI’s design is the oldest of the three and the most explicit. Two mechanisms, configurable separately:
Saved memories. Structured facts ChatGPT extracts and stores during conversation. When you say “I’m a CPA in Massachusetts working on three nonprofit audits this month,” ChatGPT may decide to store “User is a CPA in Massachusetts.” You can see every saved memory in Settings → Personalization → Memory, edit them inline, or delete them individually. The list is short and human-readable.
Chat history reference. A separate, broader mechanism where ChatGPT can recall context from prior conversations without you explicitly saving anything. This is the “wait, how did it know that?” channel. You can turn it off entirely in Settings, but you cannot see or edit individual references — it’s opaque.
The deletion model. Two distinct deletes. To remove a saved memory, you say “forget that” or delete it manually in Settings. To break a chat history reference, you have to either delete the underlying conversation entirely or turn off chat history reference globally. There’s no equivalent of “forget that one thing you remembered from last Tuesday” — it’s either everything or specific structured memories.
The opt-in. Memory is on by default for paid plans, opt-in for some regions due to local privacy regulations. There’s a memory icon in the chat composer that shows when ChatGPT just updated a memory — that’s the only indication that a memory write happened.
Best use case: Long-running personal assistant work where you want explicit control over what AI knows about you and the rest is conversational. ChatGPT’s design assumes you’ll occasionally audit the memory list and prune it. People who don’t audit end up with stale or wrong memories quietly poisoning their answers for months.
How Gemini Spark’s memory works
Google’s design is structural rather than feature-based. There isn’t a “Memory” toggle in Gemini settings the way there is in ChatGPT. Instead, Spark inherits your entire Google account as its memory layer.
What that means in practice:
- When you ask Spark to “find a coffee shop near my next meeting,” it reads your Calendar to know what your next meeting is and where, then your Maps history to know which area you prefer.
- When you ask Spark to “draft a follow-up to that client about the proposal,” it reads your Sent mail to know who the client is and what proposal you’re talking about.
- When you set up a daily Spark task to “summarize my morning emails,” it persists across sessions because the underlying state lives in your Gmail account, not in a separate memory layer.
The memory isn’t built up over time — it’s already there, in the products you’ve been using for years. Spark is a new lens on existing data.
The control surface. You manage what Spark can read through the per-app connection toggles in the Gemini app (Gmail on/off, Calendar on/off, Drive on/off, etc.). You don’t manage individual memories because there are no individual memories — Spark reads the source data fresh each session. Turning off Gmail connection means Spark instantly stops knowing about your emails, with no separate “forget my emails” step.
Persistence vs ephemerality. What Spark generates during a session (drafts, search results, plans) lives in the chat history and isn’t auto-permanent. What Spark acts on (booking a calendar entry, drafting an email reply) becomes part of your Google account’s real state. The distinction: there’s no separate “Spark memory” object — there’s just your Google account, and Spark’s actions write to it.
Best use case: Heavy Google Workspace users who already trust Google with their work data and want an agent that’s omniscient within that boundary. Worst case: people who use Google Account for personal stuff but don’t want their work AI to see their personal Gmail history. The tenant isolation is per-account, not per-context.
Side-by-side comparison
| Dimension | Claude Dreaming | ChatGPT Memory | Gemini Spark |
|---|---|---|---|
| Mental model | Offline batch curation | Explicit structured + implicit broad reference | Unified Google account context |
| When memory updates | Between sessions (background job) | Continuously during chat | Real-time from source apps |
| What gets stored | Curated workflow/error patterns | Saved facts + chat history pull | Nothing new — already in your account |
| User visibility | High (review mode, separate store) | Medium (structured memories visible, references opaque) | Indirect (per-app toggles, not per-memory) |
| Deletion granularity | High (per-entry review) | High for saved memories, low for references | Per-app only (not per-fact) |
| Cross-session continuity | Strong (post-dream) | Strong (live) | Strong (account-state) |
| Multi-user / team support | Yes, native (Managed Agents) | Limited (per-user only) | Yes, via Workspace |
| Privacy gate | Tenant isolation + review mode | Settings opt-out per-feature | Per-app connection toggle |
| Access cost | Research preview, request access | $20/mo ChatGPT Plus and up | $99.99/mo AI Ultra (Spark itself); free for account-context Gemini |
When each one wins
Use Claude Dreaming if:
- You’re running agentic workflows (not just chat) where the agent repeats similar tasks
- You want explicit control over what gets memorized and the ability to review before promote
- You’re in a regulated industry where audit trails matter — the separate memory store + review mode is the cleanest audit path of the three
- You’re building multi-agent systems where shared coordination memory matters
Use ChatGPT Memory if:
- Your work is primarily conversational chat, not multi-step agentic
- You want the most-mature implementation (this has been live since 2024 and has the longest track record)
- You’re comfortable auditing your saved memories periodically (this is a real maintenance task)
- You’re already paying for ChatGPT Plus and don’t want to add another tool
Use Gemini Spark if:
- You live in Google Workspace and your work data is already there
- You want the agent to “just know” what’s in your account without you teaching it
- You’re willing to pay $99.99/mo for AI Ultra
- You’re comfortable with the all-or-nothing per-app connection model (vs the granular per-memory edits in ChatGPT)
What each system can’t do (yet)
Claude Dreaming can’t:
- Run on individual user chat memory (it’s agent-focused, not personal-assistant-focused)
- Update in real time during a conversation (it’s a batch process)
- Replace per-conversation context — you still need the regular Claude context window
- Work without enterprise access — most individual users won’t see it in their Claude.ai chat anytime soon
ChatGPT Memory can’t:
- Run a curation pass over stale memories — you have to manually prune
- Surface why it pulled a chat-history reference (the implicit channel is opaque)
- Share memory across team members (it’s per-account)
- Distinguish “forget this fact for the next conversation” from “forget this fact forever”
Gemini Spark memory can’t:
- Remember things outside your Google account (no manual “save this fact” mechanism)
- Forget individual facts without disconnecting an entire app
- Work for users who don’t use Google products
- Be audited at the per-memory level the way Claude or ChatGPT can
What this means for you
If you’re a solo professional choosing one AI for daily work: ChatGPT Memory remains the most mature consumer-grade implementation. Use it, audit your saved memories quarterly, turn off chat history reference if you find the implicit channel annoying.
If you’re a developer building agentic workflows: Claude Dreaming is the only one of the three designed for your use case. Request research preview access, use review mode at first, watch for the same Harvey-style 6× pattern in your own agents.
If you’re a Google Workspace power user: Spark’s account-context model is the natural fit. Connect apps one at a time, test each one for a week, decide whether the convenience is worth the broader data surface.
If you’re a knowledge worker who uses all three AIs daily: Don’t replicate the same context across all three. Pick one as your “primary memory home” and let the others stay context-light. The maintenance burden of keeping three separate AI memories accurate is higher than the productivity benefit of cross-AI continuity.
If you’re an enterprise IT or compliance person evaluating these: Claude Dreaming’s separate-store + review-mode + tenant-isolation model is the cleanest audit path. ChatGPT and Spark both work for enterprise but require more policy work to get to the same level of explicit control.
The bottom line
Memory in AI used to be a research problem. As of this week, it’s a product problem with three competing answers — and each answer reveals what the lab thinks the future of AI assistance looks like. Anthropic thinks the future is agentic workflows that learn between runs. OpenAI thinks the future is a personal assistant you train explicitly. Google thinks the future is a layer on top of the data infrastructure you already have.
None of these is universally right. The right answer for you depends on where your work and data already live. The one mistake to avoid: assuming all three “memory” labels mean the same thing. They don’t, and choosing based on the marketing copy alone will leave you with the wrong system enabled in the wrong context.
If you want a structured framework for setting up AI memory across multiple platforms — what to enable, what to audit, what to delete — our Claude Code Mastery course covers the agent-memory patterns and the privacy-aware setup that applies to all three systems above. First 2 lessons free.
Sources
- Anthropic: New in Claude Managed Agents (Dreaming announcement)
- Fellow AI: What Is Claude Dreaming? Anthropic’s New AI Memory Feature
- ZDNET: Your Claude agents can dream now — how Anthropic’s new feature works
- VentureBeat: Anthropic introduces “dreaming,” a system that lets AI agents learn from their own mistakes
- Let’s Data Science: Anthropic Launches Dreaming for Claude Agents at Code with Claude 2026
- Forbes: Claude’s New Dreaming Feature Builds Self-Improving AI Agents
- MindStudio: Code with Claude 2026 — 5 New Agent Features Anthropic Just Shipped
- OpenAI: ChatGPT Memory Help Center
- The Verge: Claude memories can be exported at any time
- Ars Technica: Anthropic’s Dreaming explained
- Google blog: Gemini Spark personal AI agent
- Google blog: Everything new in Google AI subscriptions from I/O 2026
- Business Insider: Anthropic positions Dreaming as enterprise automation
- Mashable: Claude memory management
