Every Major AI Now Has Memory. Here's What Changed — and What Hasn't.

In 2026, memory became table stakes for AI assistants. Competition, advances in context engineering, and user demand to stop re-explaining themselves drove five major platforms to ship it within months of each other. The result is better recall inside each product — but your context is now fragmented across five separate silos, none of which talks to the others. Portability and ownership are the layer that hasn't shipped.

The timeline is dense and worth reading in order.

Claude — March 2026. Anthropic added memory to Claude on all plans, including free. It works by summarizing past conversations and carrying that context forward. The approach is deliberately lightweight — focused on summary-level continuity rather than granular fact tracking — and was the first major cross-tier rollout of the year, as of June 2026.

Microsoft 365 Copilot — January–May 2026. Microsoft shipped memory for Copilot between January and May 2026, reaching general availability in that window. It captures preferences, working style, and recurring topics automatically, while also letting you say "remember this" explicitly. A "memory updated" signal confirms saves; Settings surfaces everything you can view, edit, delete, or turn off. Governance controls make it enterprise-safe, as of June 2026.

Grok — May 18, 2026. xAI announced cross-conversation memory for Grok, available on web, iOS, and Android, alongside Grok 4.3. The standout addition was Skills: persistent, reusable expertise — formatting rules, workflow preferences, document styles — that Grok applies automatically without you repeating yourself each session. Not available in the EU or UK as of June 2026 due to GDPR.

ChatGPT Dreaming — June 4, 2026. OpenAI launched Dreaming, a background process that synthesizes memory from your chat history rather than only saving facts you explicitly state. It updates entries as time passes — revising "going to Singapore in July" to "went to Singapore in July 2026" after the trip ends. A roughly 5x compute reduction brought memory to free accounts for the first time and doubled storage for Plus and Pro users. The focus: staleness, correctness, and scale, as of June 2026.

Gemini. Google's memory layer for Gemini lets you save context with a "remember" prompt and review it in Settings. The feature it previously called "past chats" was renamed "memories." Personal Intelligence connects Gemini to your Google apps but is off by default; Google states it does not train on Gmail or Photos data, as of June 2026.

Competition collapsed the timeline. When one major assistant ships memory and users notice that it stops asking who they are, rivals face immediate pressure. The March–June 2026 sprint reflects that dynamic. Memory was no longer a premium differentiator — it became the baseline expectation, and every team had to match it.

Context engineering matured enough to do it cheaply. OpenAI's Dreaming is partly a story about efficiency: a 5x compute reduction didn't just make memory possible for free users, it made it practical at hundreds of millions of accounts. Claude's summary-based approach and Grok's Skills abstraction show different teams arriving at different cost-quality tradeoffs — but all of them finding the problem tractable in a way that wasn't true two years ago.

Users were paying a real productivity tax. The memory race addressed a concrete pain: you'd describe your role, your preferences, your project context — and then do it again in a new session, or in a different tool. Platform memory fixes that within a single product. That's a genuine improvement worth acknowledging, and each of these launches delivers real value on its own terms.

*The competition is now about whose memory you build.* Better recall inside an assistant is also more compelling retention engineering. The more accurately a platform models your habits and history, the higher the cost of moving to something else. Memory improves the product and builds a switching barrier — those aren't separate effects.

The wave of memory launches solved recall inside each product. It didn't solve anything beyond a single product's walls.

If you use ChatGPT, Claude, a Copilot-powered workflow, and a coding tool, you now have four or five separate, partial models of yourself. None of them reads any of the others. You still re-explain your preferences when crossing from one to another. You still can't export what one platform knows about you and import it somewhere else in a structured way. And you have no shared version history across tools — just isolated snapshots per vendor.

Grok's Skills feature shows the right instinct: persistent, reusable expertise that auto-applies is exactly what cross-tool workflows need. But Skills lives in Grok. The same logic applies to Dreaming's self-curating entries, Claude's summaries, and Copilot's preference store. Each is excellent within its context; none travels.

There's also an ownership question that the announcements mostly sidestep. Platform memory lives on the vendor's infrastructure. You can view and delete it, but you don't control the underlying store, the versioning, or the portability. As these systems get richer and more accurate, the gap between "I can edit this" and "I own this" becomes more meaningful.

The launches above are reasons to want a cross-AI memory layer more, not less. Each product getting better at knowing you individually makes it more valuable to have a single source of truth that all of them can read — rather than five improving but disconnected stores.

MemoryLake is that layer. It stores your context once, in infrastructure you own (AES-256 encryption, exportable, deletable), and makes it available to ChatGPT, Claude, Gemini-adjacent workflows, and any tool that speaks MCP through the MCP integration. Version control means you track how your context changes over time — not just what it currently says. And on the LoCoMo long-conversation benchmark, MemoryLake scored first at 94.03% as of June 2026, which speaks to the quality of retrieval when context depth actually matters. See all integrations for the current list of connected tools.

The point isn't to compete with ChatGPT Dreaming or Grok's Skills — it's to complement them. Let each platform's memory do what it does well inside that product. The cross-AI layer handles what none of them can: making your context portable, owned, and consistent across wherever you work.

Every platform knowing you better is genuinely good. The piece still missing is the single layer that travels between all of them — that you own, can audit, and can carry from one AI to the next without re-explaining yourself from scratch.

Sources: OpenAI Dreaming announcement (June 4, 2026); Anthropic Claude memory, all plans including free (March 2026); xAI Grok cross-conversation memory and Skills, Grok 4.3 (May 18, 2026); Google Gemini memory layer; Microsoft 365 Copilot Memory GA (January–May 2026). All feature details current as of June 2026; verify against each vendor's latest documentation before relying on specifics.