Not everything an AI sees should become memory.
The defining name for the control layer that decides what agents are allowed to remember.
The domain MemoryAdmission.com is for sale.
The thesis
Persistent memory is what turns a language model into an agent that improves with use. It is also a new attack surface with unusual properties: a single poisoned write can influence hundreds of future sessions, long after the conversation that planted it has ended. Published research has already demonstrated the mechanics. The MINJA memory-injection study (2025) reported poisoning an agent's long-term memory through ordinary queries, with injection success above 95 percent on the agents tested. OWASP now tracks memory poisoning as a top agentic risk, listed as ASI06 in its 2026 agentic security taxonomy.
The industry response so far concentrates on reading: retrieval filtering, output guardrails, context sanitization. The write path is comparatively unguarded. Yet every durable-state system that came before (databases, caches, orchestrators) eventually grew an explicit gate on writes. Kubernetes calls its version admission control, and the name transferred cleanly to a whole discipline of policy enforcement.
Memory admission is that idea applied to AI memory: a checkpoint where information earns its way into persistence. The term is descriptive, it borrows a respected lineage, and no one owns it yet.
Memory is shipping everywhere at once. Assistants, coding agents, and enterprise copilots now persist user facts, preferences, and learned procedures across sessions. Every one of those write paths is currently an open door.
The attacks are documented, not hypothetical. MINJA-style memory injection, delayed-trigger poisoning, and gradual goal-drift attacks all appear in peer-reviewed work from 2025 and 2026. The distinguishing property is patience: the payload waits in storage until a later task triggers it.
Privacy law meets machine memory. When an assistant remembers something it should not (health details, a colleague's data, content a user asked to keep out), the liability lands at write time. Admission policy is where consent and retention rules become enforceable.
Memory quality is a product problem too. Beyond security, unmoderated memory degrades agents: stale facts, contradictory observations, and low-value trivia accumulate until retrieval surfaces noise. Write-time curation is the fix the database world would recognize.
Sources: Memory Injection Attacks on LLM Agents via Query-Only Interaction (MINJA), arXiv:2503.03704 · OWASP GenAI Security Project · Kubernetes: Admission Controllers
Who is likely to care: Security teams threat-modeling agents that persist state across sessions; aI platform engineers who own a memory subsystem and its write path; privacy and compliance officers mapping consent onto machine memory; researchers and founders commercializing memory-integrity work.
What could be built here
Plausible directions for the category, presented as opportunities rather than existing products. Any one could anchor a company; several could coexist as a product line.
- AI memory write-gating
- An enforcement point on the memory write path: every candidate memory is scored and policy-checked before it persists, and rejections are logged.
- Long-term memory policy engine
- Declarative rules for what may be remembered, for how long, from which sources, evaluated at admission time, when enforcement is cheap and clean.
- Context poisoning prevention
- Provenance and trust scoring for candidate memories, built to catch injection payloads that are designed to look like ordinary observations.
- Enterprise memory governance
- The audit and control plane compliance teams will ask for: who taught the agent this, when, under which policy, and how do we purge it.
- Agent memory retention controls
- TTLs, consent flags, and legal-hold semantics attached at admission, so downstream deletion and discovery become tractable.
- Trusted memory ingestion API
- A drop-in admission layer in front of existing memory stores and frameworks, so teams add governance without replacing their stack.
The admission gate
Evaluation happens once, at write time. Everything downstream inherits the guarantee.
Incoming context
conversation, tools, retrieved documents
Candidate memory
extracted fact, preference, or procedure
Admission evaluation
trust, provenance, relevance, consent
Memory admission
verdict logged with lineage
- Long-term memory
- Working memory only
- Rejected
Why this name
The name states the thesis. Memory admission asserts, in two words, that memory is a privilege granted rather than a side effect. A category name that argues for the category is doing unusual work.
A lineage engineers already respect. Admission control is established vocabulary from Kubernetes and networking. The borrowed term tells an infrastructure engineer exactly where this layer sits and how it behaves.
Write-side positioning is open. Guardrails, retrieval filters, and evaluation tools crowd the read side. The write side of AI memory has documented attacks, published defenses, and no established brand.
Two words, exact match, .com. Short enough to say in a meeting, specific enough to search, credible enough for an enterprise security buyer. Domains meeting all three criteria in a live category are scarce.
Room for a product and a practice. The name supports a commercial product, an open-source admission controller for memory frameworks, or the title of the discipline itself, the way admission control named both.
Who it fits
Buyer classes, not named companies. If your team fits one of these descriptions, the inquiry form below reaches the decision-maker directly.
AI security vendors
Companies extending injection-defense and runtime-protection products to the persistence layer, where attacks now wait.
Memory infrastructure companies
Teams building agent memory layers and stores, for whom trust and governance is the enterprise differentiator.
Enterprise AI governance platforms
Vendors whose customers must answer auditors' questions about what their agents have learned and from whom.
Model and platform providers
Organizations shipping first-party memory features who need a name for its safety layer.
Research groups commercializing
Labs behind the memory-integrity literature, spinning defenses into products that need a definitive banner.
Reading
What is memory admission?
An explainer on memory admission control for AI systems: why durable agent memory needs a write-time policy gate, what the research shows, and how the pattern works.
Buying it
No price is listed. A credible intended use, and an indicative range if you have one, move the conversation fastest.
How a sale works
- Transfer via a reputable escrow service or registrar-mediated process. Funds release only on confirmed transfer.
- Confidential discussions welcome; a mutual NDA is available before substance.
- The sale conveys the domain registration. Trademark clearance for your use is yours to run, as with any name.
Questions
Yes. MemoryAdmission.com is available for acquisition by a party building in AI memory, security, or governance.
No. This site presents the strategic case for the domain name itself. No product or service currently operates under this name.
The sale covers the domain name registration only, unless otherwise agreed in writing. No trademark rights are claimed or conveyed; buyers should perform their own trademark clearance for their intended use.
Through a reputable escrow service or registrar-mediated transfer. Funds sit in escrow, the domain moves to your registrar account, and funds release on confirmation, typically a few business days end to end.
Yes. Inquiries are treated as confidential by default, and a mutual NDA is available before substantive discussion.
No fixed price is listed. Including an indicative range with your inquiry is optional and tends to shorten the conversation.