The Short Answer
A K-12 AI agent should keep the durable facts a student needs honored, such as screen-reader support or a preferred language, and let temporary state expire. ibl.ai's long-term memory toolkit, shipped 11 September 2026, scopes memories to one agent by default, attaches an expiry to temporary facts, and purges expired rows nightly. With ibl.ai you own all the code and the data, so retention is district policy, not a vendor setting.
Most district conversations about agent memory are about what the agent should remember. The harder control, and the one that maps onto student privacy law, is what it should forget.
What did ibl.ai ship in the long-term memory toolkit, and when?
A toolkit that lets an agent manage its own long-term memory, released in the platform update covering 2026-09-04 through 2026-09-11.
Graph agents running on tool-calling models now save, update, forget and search memories directly, instead of a separate worker LLM call running after every turn.
Three properties matter for a district. Memories default to the agent's own scope. Only high-level facts — name, role, language, accessibility needs, broad goals — are stored globally. Temporary state carries an expiry.
Expiry is enforced by a scheduled task, purge_expired_memories, which runs nightly at 04:20 UTC. The database migration that carries it adds a nullable, indexed expires_at column with no backfill.
DISABLE_MEMORY_EXTRACTION is a single kill switch for both the toolkit and the worker. A district that wants agents with no long-term memory at all sets one variable.
That release shipped alongside 61 others in the same seven-day window — 62 production releases in total. The full changelog is at the September 11 platform update.
One correction to how this is being described. The dual-layer content moderation that K-12 districts run is not part of this release. It is an existing platform capability, and it is worth keeping the two separate: moderation decides what an agent may say, memory decides what it may keep.
Why should an AI agent in a K-12 district forget things on purpose?
Because the two kinds of things an agent learns about a student have opposite half-lives.
Some facts should persist and be honored every time. A student uses a screen reader. A student's family communicates in Portuguese.
A student has an accommodation that changes how material should be presented. An agent that forgets those makes the student re-declare a need at the start of every session.
Other facts are episodes. Failed the Unit 3 retake. Struggling with the current unit. Sitting an exam this week. These are true for a few weeks and then they are not.
An agent that keeps both categories forever produces a specific failure: a bad fortnight in seventh grade follows a student into ninth. It is what a store with no expiry does by default.
This is also a pedagogical point, not only a privacy one. A system that silently conditions every future interaction on a student's worst month is not neutral about that student's prospects.
How does memory expiry map onto FERPA and COPPA retention expectations?
It turns a retention policy into something the software executes on a schedule rather than something a district asserts in a data protection agreement.
The FTC's amended children's privacy rule is explicit on the point.
Operators must "only retain personal information for as long as reasonably necessary to fulfill a specific purpose for which it was collected," and the provision "explicitly states that operators cannot retain the information indefinitely."
The Commission announced the final rule on January 16, 2025, and the amendments were published in the Federal Register on April 22, 2025. Full compliance has been required since 22 April 2026, so this is an operative rule, not a forthcoming one.
The operator obligation in the FTC's own compliance guidance reads the same way: retain a child's personal information "for only as long as is necessary to fulfill the purpose for which it was collected and delete the information using reasonable measures," per the COPPA FAQ.
FERPA works differently and is often misread here. It does not impose a general destruction deadline on education records, and many states require long retention of attendance and graduation records.
It does set one bounded exception in the other direction: 34 CFR § 99.10(e) bars destroying education records while a request to inspect and review them is outstanding.
What the Department of Education's Privacy Technical Assistance Center recommends is data minimization — destroying data once it is no longer needed — in its Best Practices for Data Destruction guide.
An expires_at column does not make a deployment compliant with either law. What it does is give a district a place to put the answer.
Retention becomes a field with a value, a nightly job that acts on it, and an audit trail, instead of a paragraph in a policy document that no system reads.
A caveat worth stating plainly: this is a platform capability. It does not describe any district's deployment, and compliance is a property of how a district configures, hosts and governs the system.
What does dual-layer content moderation cover in a K-12 deployment?
Two screens, one on the way in and one on the way out.
Every student message is screened before it reaches the model, and every model response is filtered before it reaches the student. The rules are configurable by grade band — K-2, 3-5, 6-8, 9-12 — so a second-grader and a high school junior get different thresholds.
Districts define what topics are off limits, what language is flagged, and how an agent responds to sensitive questions. Every interaction is logged and auditable.
Moderation and memory are separate controls that are easy to conflate.
Moderation governs the content of a turn. Memory governs what survives the turn. A district needs both configured, and configured independently — a message can be perfectly appropriate to answer and still be something you do not want retained for two years.
The K-12 roster is 12 agents, from tutoring and lesson planning to special education and family communication, listed at ibl.ai's K-12 solutions page.
What should a district CTO ask before turning agent memory on?
Five questions, all answerable from configuration rather than from a vendor's assurances.
- What is stored globally, and who decided the list? Accessibility needs and language belong there. Performance episodes do not.
- What carries an expiry, and how long is it? If the answer is "nothing," the retention policy is "forever."
- What runs the purge, and can you see it run? A scheduled task with a name and a log line is auditable. A promise is not.
- Can a student or guardian request deletion, and does it reach the memory store? A deletion that clears chat history but leaves extracted memories is not a deletion.
- Can you turn the whole mechanism off? A single documented kill switch is worth more than a settings page, because it is testable.
The related question — what memory and skills actually are inside an agent loop — is covered in memory and skills: what turns an agent loop into a real AI agent.
How does ibl.ai deploy agent memory for school districts?
Inside the district's own perimeter, where the retention decision is the district's to make.
With ibl.ai you own all the code and the data.
The platform runs on district infrastructure with full source code access, is model-agnostic across any LLM, is usage-based with no per-seat pricing, and can deploy anywhere — your own cloud, on-premise, GovCloud, or a fully air-gapped network.
A district that owns the code can read the purge task, change the expiry defaults, and prove to a school board what the system keeps.
That inspectability is the practical difference. A hosted product can tell you its retention policy. A platform you run lets you read the migration that added the column and the task that empties it.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY.
Related reading: COPPA compliant AI for schools — where under-13 student data sits during the inference call; and memory and skills: what turns an agent loop into a real AI agent — what memory is doing inside the loop in the first place.
Sources: the memory toolkit, expiry, nightly purge_expired_memories task and kill switch from ibl.ai's September 11, 2026 platform update; the COPPA retention provision from the FTC's January 16, 2025 final-rule announcement and the FTC's COPPA FAQ; the FERPA data-minimization guidance from the Department of Education's Best Practices for Data Destruction; the outstanding-request bar from 34 CFR § 99.10(e).