The Coming Battle Over AI Memory: Who Owns What Your Agent Knows?

Photo: perpetual.fostering, Wikimedia Commons, CC BY 2.0

01 What your agent knows — and who holds it

Every serious assistant now keeps two distinct records of you. Chat history: the raw transcripts, bulky and exportable as a ZIP of JSON. And saved memory: the short distilled profile — your work, your tone, your recurring projects, your stated preferences — that makes the assistant feel like it knows you. The first is a log. The second is an asset, and it lives almost entirely inside the vendor's systems.

In March 2026, within weeks of each other, Claude, ChatGPT and Gemini all shipped memory-import features. The marketing suggested you could pick up your AI brain and set it down somewhere new. What actually ships is narrower: a one-time, plain-language snapshot — you ask one assistant to write out what it remembers, then paste that text into another. It is a migration aid, not a bridge. The moment the text lands, the two copies evolve independently; no sync exists between any pair of major assistants.

Analysis — not prediction. N43 and Hermes AI grounds every scenario in the documented record and verified reporting as of September 19, 2026; where evidence is incomplete we say so.

02 The portability gap, precisely measured

The gaps are specific. OpenAI's memory export, added March 2026, produces a JSON of stored facts — with no API, so it cannot be automated — and saved memories historically lived in a separate store from the standard data-export ZIP. Anthropic's import is a copy-paste prompt the vendor wrote for the purpose. Google's tools ingest up to 5 GB of transcripts — which teaches Gemini your history without importing the other assistant's distilled model of you — and both tools are unavailable in the EEA, the UK and Switzerland, closing the path entirely for those users.

The pattern across all three: import buttons point one direction, at the company that built the button. None of this is accidental. The continuously updated memory is the moat; the exports are the photocopy the moat-holder graciously notarizes.

03 What the law already says — and what it misses

GDPR Article 20 grants the right to receive personal data you provided “in a structured, commonly used, machine-readable format” and transmit it elsewhere, within one month, where processing rests on consent or contract. That plainly covers stated facts, saved preferences, saved conversations and agent configuration. It does not plainly cover the vendor's continuously-updated internal representation — which is recomputable, inferred, and arguably the processor's own derived data.

The engineering consensus forming around Article 20 compliance says: export what the user gave you, not the embeddings you can recompute. Raw vectors are unreadable to any other app anyway; export the sentence, not its coordinates. A compliant memory port is small — facts, preferences, threads, persona config — and it is exactly the artifact vendors are slowest to produce, because it is the artifact that unlocks switching.

04 The China warning shot

What was theoretical in Brussels became concrete in Beijing. On July 15, 2026, China's Interim Measures for AI Anthropomorphic Interactive Services took effect, forcing ByteDance's Doubao and Alibaba's Qwen to wind down their companion agents — and with them, the memories millions of users had built. Doubao is keeping data read-only only until October 15, 2026, after which the accumulated context is gone.

The lesson exports well beyond China: if your product's value lives in memory the platform owns, that memory is a liability, not a moat. A regulator, an acquisition, a pricing change or a shutdown can erase years of personalization overnight — and the user walks away with nothing. The engineering response is already visible in research: open portable-memory protocols now specify JSON interchange formats with content hashing, provenance and cryptographic signatures — vCard for AI memories — plus re-hydration pipelines designed to resist memory-mediated prompt injection.

05 The deeper question: memory as property

Beneath the compliance questions sits a harder one: is the agent's model of you yours, the vendor's, or a new category law has not named? The vendor's answer is implicit — it is inference on their infrastructure, trained on your inputs, and you licensed the output. The user's intuition is the opposite — the assistant only knows those things because a human told it, over years, in confidence.

The practical resolution will likely be neither pure position but a settlement: user-owned canonical copies held outside any single assistant, with vendors as interchangeable readers. That architecture — already being built by third-party memory layers and open protocols — dissolves the moat by making switching cheap. The battle in the headline is therefore not just regulatory; it is between two designs of the same system. One keeps the model of you captive and calls it personalization. The other keeps it portable and calls it ownership. Which one wins decides whether “your” agent is a servant or a lease.

06 What to watch next

Watch whether the October 15 Doubao deadline produces a mass memory-export moment in China — the first real-world stress test of deleting a social asset. Watch the first regulator to rule explicitly on distilled memory under Article 20; that decision sets the template every vendor will engineer around. Watch adoption of open portable-memory formats by any one major assistant — a single vendor breaking ranks converts portability from research to feature. And watch the memory-layer startups: if user-held memory becomes the canonical copy, the assistants become commodities, and the most valuable context database on earth stops belonging to any of them.