I used Gemini’s new AI memory importing feature, and now it knows as much about me as ChatGPT

What happens when switching assistants is no longer a restart but a handover of a digital life!

A cursor blinks, a file uploads, and a chatbot that once greeted like a polite stranger now finishes sentences with the familiarity of a long-serving assistant. The moment the import completed, a previously forgotten travel plan resurfaced in a Gemini reply, then a project timeline appeared without prompting. It felt less like switching apps and more like moving into a home where the previous tenant left a perfectly organized closet.

On the surface this reads as simple user convenience: no more copy and paste, no more rebuilding six months of preferences. That is the dominant storyline many outlets have run with. The quieter industry shift is that memory import features are reworking the economics and legal calculus of AI platforms in ways that matter far more to businesses than to hobbyists, and that change is only now coming into focus.

Why press releases and product reviews are the primary sources here

Coverage from product tests and vendor materials has driven most reporting about memory imports, and much of this article leans on those press and hands-on accounts to explain technical flows and policy claims. That matters because the public record is often the only visible trace of how vendors treat imported data and whether it feeds model training.

The obvious reading: portability as a user-friendly win

For users, migrating context between assistants stops the digital version of losing one’s keys. Gemini’s new Import AI chats feature makes it possible to upload exported conversation files so prior threads and preferences show up inside the assistant, reducing the friction of changing services. According to Tom’s Guide, the flow is straightforward and surprisingly smooth based on early testing. (tomsguide.com)

The underreported angle that should worry product leaders

The hard consequence is that portability can be a lever for platform control or data capture. Some vendors import distilled memories while others ingest full conversation logs into platform activity stores, and that distinction changes who gets to repackage user knowledge for model improvements or commercial reuse. Resultsense reported that Google’s approach tests importing full chat histories into Gemini Activity, which may be used for model training, while competitors like Anthropic emphasize encrypted memories that are not used for training. (resultsense.com)

Competitors, timing, and why now matters

OpenAI, Anthropic, Google, Microsoft, and smaller players are all jockeying for the assistant layer, where stickiness is measured in how much context an AI retains about a user. Anthropic expanded Claude’s memory and added an import flow in early March 2026 to lower switching costs, leaning into privacy-forward messaging. Axios documented Anthropic’s memory rollout last October and subsequent moves into import tools designed to ease migration. (axios.com)

The technical plumbing behind importing memories

Exports typically arrive as zipped archives containing conversations.json or activity logs, which are then parsed and normalized into a memory store. Community projects and standards work have already begun to codify how that conversion should happen, with a Portable AI Memory specification proposing a Normalized Conversation Format and importer versioning to avoid silent breakage. That spec lays out a pipeline from raw provider export to a normalized memory store for reuse. (portable-ai-memory.org)

What actually moves and what gets left behind

Not every export transfers everything. Most import flows move profile details, preferences, and distilled project notes but exclude provider-specific features, file attachments, and custom tooling. The practical result is a partial handover of “who you are” and “what you care about” while preserving platform-specific assets where vendors prefer to retain lock-in. That nuance is the reason savvy legal teams will pause before greenlighting mass migrations.

The cost nobody is calculating for business migrations

A midmarket software team that switches primary assistants could save an hour per person the first week by importing memories, yet the hidden cost shows up in compliance and model risk. If imported logs are ingested into a vendor’s training pipeline, that company could surface corporate strategy or customer data back through model outputs unless contractual controls stop it. The math is simple: 50 employees at 1 hour saved each week equals 50 hours of regained productivity, but the potential exposure from a single sensitive prompt appearing in a public-facing output could cost orders of magnitude more in legal and reputational damage.

Memory portability turns vendor switching from a consumer convenience into a corporate governance problem.

Practical scenarios businesses should run now

An agency moving client workflows from ChatGPT to Gemini must export conversations.json from ChatGPT, audit the archive for client confidentiality, and then decide whether to redact before importing. Decrypt’s user guide shows how to export ChatGPT data, and it highlights that the archive contains full message metadata and shared context that can be portable if left intact. (decrypt.co)

A safer, pragmatic approach is to run a two-stage import: first transfer only nonconfidential profile memories, then selectively add project threads after legal review. That workflow costs time up front but reduces downstream risk, and tooling based on the Portable AI Memory spec can automate provenance tagging so imports are auditable. (portable-ai-memory.org)

Risks, governance gaps, and the slippery slope

Regulatory exposure, inadvertent training of vendor models, and unclear deletion semantics are the three biggest immediate risks. Vendors’ claims about encryption or nontraining are often framed in product pages, but independent verification is rare and difficult. Moreover, a business that allows employees to import client conversations into a vendor system without strict policies is effectively outsourcing parts of its data governance.

What product teams and CIOs must ask vendors today

Demand explicit SLA language about whether imported memories are used for model training and insist on audit logs showing when imported artifacts were accessed. Require an export-import dry run on a nonproduction dataset and check that importer versioning preserves timestamps and authorship. If a vendor resists these requests, that should be a red flag for enterprise procurement.

Where this leads the industry in the next 12 to 24 months

Expect a bifurcation: privacy-first vendors will market nontraining, encrypted memory stores as an enterprise advantage, while more vertically integrated cloud vendors will push full-history imports that accelerate model improvements and product personalization. Standards groups will gain influence, and legal frameworks for consented memory portability will emerge as an explicit procurement item for regulated industries.

The practical implication is straightforward: treat memory imports like data migrations, not feature toggles, and add them to vendor risk assessments.

Key Takeaways

• Memory import features remove the practical cost of switching assistants, turning migration into a policy decision for businesses.
• Not all imports are equal; some vendors import distilled memories while others ingest full chat histories that may feed model training.
• Enterprises must audit exported archives and require contractual guarantees about training use and deletion semantics.
• Portable standards and provenance tagging will become essential controls for safe AI migrations.

Frequently Asked Questions

Can I export my ChatGPT data and move it into Gemini without risk?
Exporting is straightforward and creates a conversations.json archive, but risk depends on what the receiving vendor does with the imported data. Audit and redact sensitive items before importing and require vendor assurances about nontraining and encryption. (decrypt.co)

Will importing memories make a chatbot smarter about my company trade secrets?
If the import includes detailed conversations and the vendor uses imported logs for training, then yes, models could indirectly surface that knowledge. Ask vendors for explicit nontraining clauses and proof of data isolation. (resultsense.com)

Are there standards that make imports interoperable and auditable?
A Portable AI Memory specification defines normalized conversation and memory schemas and recommends importer versioning to preserve provenance, which helps make imports auditable and reversible. (portable-ai-memory.org)

How should legal teams treat memory imports in vendor contracts?
Treat imports as data transfers subject to the same controls as other sensitive data: require processing agreements, deletion guarantees, and penalties for unauthorized reuse. Consider including audit rights and independent verification. (axios.com)

What immediate steps should an IT leader take if a vendor offers import tools?
Run a test import with sanitized data, confirm retention and training policies, and update acceptable use policies to prevent accidental exposure. If the vendor cannot guarantee nontraining or auditable deletion, delay production imports.

Related Coverage

Readers who want to go deeper should explore stories about AI data governance, model training consent frameworks, and enterprise procurement for AI assistants. Coverage of vendor lock-in strategies and how cloud providers monetize user data will also help decision makers build safer migration playbooks on The AI Era News.

SOURCES: https://www.tomsguide.com/ai/google-gemini-now-lets-you-switch-chatbots-without-losing-everything-i-tried-it, https://www.axios.com/2025/10/23/anthropic-claude-memory-subscribers, https://portable-ai-memory.org/spec/v1.0/, https://www.resultsense.com/news/2026-03-03-anthropic-launches-memory-import-to-ease-ai-switching, https://decrypt.co/359976