Last month, Meta, Google, Microsoft, Databricks, Couchbase, Cohere, and GitHub all rolled out new memory capabilities for AI agents.
We rounded up the biggest updates and what they signal for the AI industry. 👇 https://x.com/i/article/2118081847045636096
EVERYONE Shipped AI Memory in October, but Who Actually Owns It?
October was the month agent memory became table stakes. Across the AI stack, Meta, Google, Microsoft, Databricks, Couchbase, Cohere, and GitHub all shipped or expanded memory.
Now, the question is shifting from "Can an agent remember?" to "Who controls what it remembers?"
At the model and agent layer
- Meta added Active Memory to Muse Wonder, letting agents carry important context across conversations and bring back relevant details when the task calls for them.
- Google's Gemini agent introduced session, semantic, procedural, and episodic memory.
- Microsoft introduced shared business-process memory for Copilot Studio agents into early access through its Frontier program.
At the data and platform layer
- Databricks launched Agent Memory on Lakebase, with a background "dreaming" pipeline that consolidates memories.
- Couchbase launched Managed Agent Memory in Capella, giving agents a hosted way to extract, store, and semantically recall information across sessions.
- Cohere's North 2 added cross-session memory to its enterprise agent stack.
Even GitHub made AI memory impossible to ignore
GitHub closed October by making Copilot Memory generally available across its coding surfaces. Its "dreaming" system curates memories, while Copilot Fusion can route work across cloud and local models. That combination shows where this is heading.
You can switch AI models, but your memory still gets left behind
Taken together, these launches point to one thing: the model is becoming interchangeable, and memory is becoming the layer that persists. But model-agnostic does not automatically mean agent- or platform-agnostic. In most of these systems, memory moves across the provider's own surfaces, not beyond the ecosystem that created it.
That distinction matters in practice. Most builders don't run one agent on one stack – you might use GitHub Copilot in your editor, Claude Code in the terminal, and Cursor for a refactor, each one learning your preferences on its own and none of them sharing what they learn. Switch tools or models, and the context that made an agent useful stays behind, and you start over. Memory you can’t take with you is just storage you’re renting from one vendor.
The future of AI memory shouldn't come with vendor lock-in
Walrus Memory is built to go beyond that boundary – the layer outside any one model, agent, or provider. Persistent, portable, and verifiable memory that moves with your agents and stays under your control.