Who Will Own Our Memories?
… but first here’s the latest AI news:
… but first here’s the latest AI news:
Textbooks are evolving. Google NotebookLM’s new Learn Your Way feature presents a vision for textbooks of the future that are dynamic and multi-modal. Instead of presenting students with a static PDF to read, textbooks are adapted to each student’s grade level and made multi-modal: incorporating narration, slides, personalised examples, and memory aides.An early evaluation and RCT indicate high quality content and evidence of learning gains.
AI regulation is entering new territory. California is close to passing a first-of-its-kind bill to regulate AI chatbots, with specific guardrails for child protection. If signed into law, it could set a precedent for regulations on student facing AI and EdTech. Similar regulations are already coming into play around the world, with Italy bringing the EU AI’s act into effect with new legislation. Across Africa, countries have been drafting national AI strategies after the 2025 Africa Declaration on Artificial Intelligence was endorsed by 49 countries, committing to building governance, data protection, and institutional capacity.
The race to build compute is heating up. OpenAI announced plans to spend $100B on backup servers over five years, while Huawei unveiled new compute infrastructure for China. Governments are catching on too: Cassava Technologies and Accenture announced a partnership to accelerate sovereign AI adoption across Africa, starting with data centers in Kenya, Nigeria, and South Africa. The United Kingdom, Sweden, UAE, and Saudi Arabia aren’t far behind either.
How persistent memory works in education
Persistent memory is made up of three fundamental concepts today:
Secure record keeper - A safe place, apart from any one app, where each learner’s progress can live day-to-day, year-to-year.
Relevance filter - Simple rules that surface only what helps this learner with today’s lesson context and hides the rest.
User controlled switches - Clear options for parents, teachers, or ministries to see, share, or erase those records whenever they choose.
Until very recently, LLM-powered bots were mostly faking it - they were remembering “what” and “how much” to store in their memory using a mix of your explicit instructions and their own “is this important?” heuristics. They had very limited storage that selectively and opaquely “added to memory” small snippets of information that you fed it over time, but often it felt incomplete or inadequate, and it was.
Here’s a great illustration of how persistent memory works today, inside a frontier lab’s LLM bot, Claude.
And another fun TikTok, comparing memories across the tools:
Persistent memory and context is the next big leap for the frontier labs. We’ve seen a rapid deployment cycle in the past few months focused on just this.
LLM-powered bots and tools will be starting to remember more very soon, but those memory layers are currently primarily proprietary; they cannot be ported over to other tools, and they are disincentivized to work together to solve this problem, like VHS vs Betamax.
To solve for this, we’re seeing open persistent memory layers, like Mem0 (mem-zero) come out. Mem0 is like a portable flash drive for AI memory - basically any tutoring bot could plug in and remember a learner’s progress without holding the data hostage. Think of it as an external, durable knowledge storage you own - a database with policies - that any LLM can read and write to using APIs. It survives model swaps, app restarts, and account changes because it’s not tied to a single chat product.
We’re still in early days, but these types of tools give an indication of what’s possible with persistent memory.
How could persistent memory impact teaching and learning?
Imagine this:
Meet a 6 year old student named Lena. She’s starting P2 this academic year in Lagos. She packs her backpack in the morning with her father, and the two of them walk together to school, in anticipation. They arrive at her classroom where Lena’s new teacher, Elvin, greets Lena and her father, and asks them about their holidays. Elvin pulls out his phone and asks Lena’s father if he still permits the school to access Lena’s “Lockbox”, her persistent memory. Lena’s Lockbox contains her core identity and attendance logs, her mastered concepts, her reading fluency, her phonetic pronunciations, her preferred style of learning, her self-reported check-ins, her frustration patterns, her misconceptions, her teacher comments, her peer reviews, and her and her guardian’s consent approvals. With permission, Elvin opens up Lena’s Lockbox and reminds Lena's father what data is available to structure Lena's learning plan. After speaking to her father about Lena's summer activities (reading 5 books), Elvin updates her Lockbox. Elvin reminds Lena’s father that he and Lena (once age appropriate) are the ultimate owners of her Lockbox, not the school nor the AI apps. They can access it, decide which apps do or do not get access, and where it lives. And the MoE can see anonymized and aggregated data about the school as well, to ensure Lena and other students are tracking properly. Lena says goodbye to her father, sits down at her desk, and Elvin begins the school day.
He starts with English. He reminds the class that to get a sense of where everyone is at after the holidays, he’ll administer a quick assessment, which he’ll feed into their Lockboxes. Then he’s going to group the class into teams based on their Lockboxes, to help him better target instruction. “Let’s begin,” Elvin says.
This picture could be a reality very soon, with the pace of innovation in AI.
What principles need to be in place to make this future a reality?
We, the funders, the governments, the vendors, and the implementers, get to guide which future prevails. We have the opportunity now, as the AI tools take shape and pick lanes, to ask together where and how and what memories we are storing about learners, what their “Lockboxes” will look like, and who will hold the key.
To guide this reality, we would suggest a focus on four areas:
Prioritize interoperability in policies / law / tenders / RFPs. Governments and funders require interoperability in their tenders and establish clear standards for use.
Support the use of public utility rails. This means identifying a neutral third-party, often owned by a combination of vendors and government, to build and maintain a public API interface.
Incentivize adoption through procurement criteria. In order for anyone to bid for government funded, or donor-funded projects, they’ll need to commit to complying with open standards on persistent memory.
Promote competition to discourage early monopolies. The goal is to make it illegal or at least implausible for vendors to lock in their consumers using proprietary memory layers that don’t transfer.
A Cautionary Tale: a Locked-in Lockbox
Now picture the same classroom, only this time Lena’s Lockbox lives on a proprietary LLM’s foreign server, one she doesn’t own and can’t see. Her AI tutor remembers every typo and key stroke, and has labeled her a “slow reader.” Her school is keen to switch to a different AI product, but the Lockboxes for all their students are held within a closed, big tech-owned cloud server, and the school and government don’t have any way to access nor transfer those memories to other tools. Elvin feels frustrated because there’s no way for him to view, let alone augment and update the Lockboxes, so he finds himself relying on the algorithms to lead the class. “What’s the point of intervening?” he asks himself. When Lena’s father comes to pick her up from school, he asks Elvin to wipe Lena’s data from the Lockbox, since they’re moving to Oshodi. Elvin is at a loss, he doesn’t have the ability to fulfil her father’s request, and Lena will have to start in Oshodi from scratch.
We’re starting to see the early choices being made that will determine which of these scenarios become our future reality. We have an opportunity to define, clarify, and ultimately streamline the rules for who will own our memories, guided by what these choices could mean for kids’ learning experiences and their learning outcomes.
This situation is not dissimilar to that of mobile money, only 10 years ago. Before 2016, Indians relied on closed prepaid wallets, like Paytm. Balances were trapped inside each app and shops had to juggle multiple wallets to support consumer demand. In 2016, the National Payments Corp of India (NCPI) rolled out a unified payments interface, an open and real-time payments rail built on standard public APIs, creating an interoperable and level playing field for all wallets, and freeing customers from single app dominance.
Had the government and funders not stepped in to restrict private companies from monopolising the market, had they not created supporting policies and open standards to encourage competition while enabling users to own their own wallets, the Indian markets and consumer experience would look very different today.
Taking action
As persistent memory features continue to evolve, ministries may need support to navigate hosting and data ownership negotiations with vendors, and funders have a role to play in providing that support. International benchmarks, such as UNESCO’s guidance on AI in education or the OECD AI Principles, could also help guide interoperability and governance.
On the infrastructure side, financing and coordination will be essential to build public utility rails that can support open, portable memory layers. At the same time, teachers and parents will need support to understand how to use this data responsibly. This could include integrating with existing interventions in education like TaRL and strengthening awareness of their rights as data owners. The choices made now will determine whether persistent memory can empower learners and educators globally.
This edition of the Learning Futures Briefing was led and written by Shabnam Aggarwal, with contributions from Dr. Robin Horn, Ayesha Khan, Sara Cohen, and Faizan Ul Haq.
