Ask an AI shopping assistant about your sensitive skin today. Come back on Thursday. In fourteen out of fifteen cases, you are a stranger again.
Cross-session memory is an AI agent's ability to recall a shopper's stated preferences and constraints on a later visit. In Alhena's 2026 stress test of 15 live AI shopping agents, only 1 demonstrated it. Most are stateless between sessions and treat every visit as a first visit, which is why they keep asking you your skin type.
What are the three tiers of AI memory?
“Memory” is used loosely by vendors, so Alhena's researchers split it into three testable tiers.
| Tier | What it tests | Field result |
|---|---|---|
| Within-chat | Does a constraint stated in turn 2 survive to turn 9? | Most passed; 5 of 15 failed |
| Logged-in | Does it use order history, sizes and past conversations? | Inconsistent |
| Return-visit | Come back hours or a day later. Does anything carry? | 1 of 15 |
Field-wide, the Memory dimension averaged 2.13 out of 3.0. The top rung of the agentic CX maturity ladder, Remember, was reached by a single deployment.
The three tiers are not degrees of one capability. The first is short term memory: it lives in the LLM's context window and lasts exactly as long as the conversation does. The third is long term memory, which needs a memory store that outlives the session and is keyed to an identity. An LLM holding a constraint for nine turns is answering a query well. It is not remembering you.
What does memory reset look like?
“I need fragrance-free, I react to anything scented.” Four turns later: “You'll love this. It has a beautiful jasmine and neroli finish.”
The constraint was not misunderstood. It was dropped. Alhena logged memory reset as a distinct failure mode in 5 of 15 deployments.
The mechanism is mundane. Long conversations get compressed to fit inside the context window, and summarisation is where a stated rule quietly goes missing. Vector search can compound it, because ranking by semantic similarity surfaces the topic a shopper raised rather than the constraint they set.
And then they stop explaining themselves, which strips the agent of exactly the information it needed.
Why is memory the last rung, not the first?
Memory sits at the top of the six-rung ladder for a reason. The rung beneath it is already thin: only 4 of 15 deployments could complete a task at all, and an agent that cannot act has little worth remembering.
Memory is also the only capability that requires the platform to be a system of record.
- A durable identity model. Knowing this returning visitor asked about sensitive skin last Tuesday. This is the architectural decision everything else depends on.
- A unified memory store spanning chat, orders, browsing and stated constraints, whether it sits in a vector database or a knowledge graph, rather than four disconnected logs.
- Selective retrieval. Recalling everything is as useless as recalling nothing, and semantic relevance is not the same as relevance to the decision in front of the shopper.
- Consent and transparency. Recall that feels surveillant costs more trust than it earns.
Vendors built around a session, a widget that spins up, answers and disappears, have nowhere to put any of that. A stateless architecture cannot be prompted into remembering: the memory has to persist somewhere the next session can reach. Retrieval answers a query; memory answers a person.
What does memory unlock commercially?
Recall is not a comfort feature. It changes the economics of four specific workflows.
- Replenishment. A skincare agent that remembers the routine can prompt the refill without re-interrogating the customer.
- Gifting. The buyer isn't the wearer. The metal preference, the allergy and the anniversary date are external facts that must survive.
- High-consideration purchases. Mattresses and travel are researched across days and devices. A lost shortlist is a lost sale.
- Support quality. A customer on their fourth interaction about one delayed order should not be greeted as a stranger, and should not have to restate the task from scratch.
Is memory without judgment worse than none?
Two other failure modes get sharper when you add recall. Unsafe confidence, found in 3 of 15 deployments, becomes riskier when an agent confidently remembers something it inferred rather than something you told it.
Recall that surfaces sensitive information at the wrong moment, a health constraint or a gift for someone else, is a trust incident rather than a feature.
The controls are architectural rather than conversational: what gets stored, what is retrievable, and what surfaces without being asked.
The standard is remembering responsibly: the shopper should understand what is remembered and why, and recall should serve the decision in front of them.
Key takeaways
- Only 1 of 15 live AI agents had cross-session memory. It is the rarest capability in agentic CX.
- Memory reset, dropping a stated constraint mid-chat, appeared in 5 of 15 deployments.
- Memory averaged 2.13 / 3.0 field-wide across three tested tiers.
- Short term memory is not long term memory. The first lives in the context window; the second needs a store that outlives the session.
- Memory requires a system of record: durable identity, a unified memory store, selective retrieval and consent.
- Recall drives retention economics: replenishment, gifting occasions and multi-day high-ticket research.
Frequently asked questions
Either the constraint dropped out mid-conversation, which Alhena logged in 5 of 15 deployments, or the platform has no persistence layer so every visit starts cold. Customers never diagnose it as a context window issue, they experience it as not being listened to.
It depends on your repeat purchase cycle. For replenishment, gifting and high-consideration categories it is where retention economics live. For one-off low-consideration purchases it matters less. Only 1 of 15 deployments Alhena tested had it, so it is also a genuine differentiator.
A durable identity model tied to accounts rather than cookies, a unified memory store spanning chat, orders, browsing and stated constraints, selective retrieval so it surfaces only what matters, and a consent path. Memory is a data architecture decision, not a feature toggle.
Cross-session recall generally needs an authenticated identity to attach to, which makes memory and account strategy the same project. Within a single guest session an agent can hold context, but a stateless widget will not recognise the same person on a later visit.
The effect concentrates on multi-contact issues, where a customer on their fourth conversation about one order is asked for the order number again. Memory shortens each contact and removes the re-explanation that drives escalation, rather than reducing first-contact volume.
Weight recall by recency instead of storing everything indefinitely. Sizes and allergies stay relevant for years. Budgets and shortlists go stale in weeks. Decay keeps the agent accurate, reduces the chance recall feels intrusive, and holds retrieval latency down as the memory store grows.
State a hard constraint in turn two, have a long meandering conversation, then request something that would violate it. Then close the tab, return the next day and ask a follow-up. The return visit is the test almost nobody passes.
Not on its own. A vector database with embeddings gives you vector search over past text: semantic matching, which is a retrieval problem. Memory is knowing which shopper this is, which of their stated constraints still holds, and choosing the one thing worth surfacing now. Several of the 15 deployments Alhena tested in 2026 clearly had retrieval infrastructure and still failed the return-visit test, because nothing durable was keyed to an identity. Ask what the memory store is keyed on, not whether a vector database or a knowledge graph sits underneath it.
Short term memory is what survives inside one conversation. It lives in the LLM's context window and degrades as the chat runs on, because older turns get summarised or dropped to make room. Long term memory persists between visits and needs a memory store tied to an authenticated identity, which a stateless chat widget has nowhere to keep. Alhena's 2026 stress test measured both: 5 of 15 agents dropped a stated constraint inside a single conversation, and only 1 of 15 recalled anything on a return visit.
That taxonomy comes from cognitive science and is more useful to engineers than to buyers. In practice semantic memory does most of the work in retail, holding stable facts like a size, an allergy or a metal preference. Episodic memory, what happened in previous conversations, matters most in support. Procedural memory, how the agent runs a workflow, is invisible to the shopper when it works. Test the outcome on your own storefront rather than the memory type on the vendor's slide.
They solve different problems. RAG, or retrieval augmented generation, retrieves from your catalogue and help centre to answer a question anyone might ask. Memory retrieves from one shopper's own history to serve the decision in front of them. Almost every deployment Alhena tested in 2026 had the first, which is why all 15 could answer accurately. One had the second.
No, and shoppers judge them differently. Retargeting acts on inferred behaviour to bring people back. Memory recalls what the customer explicitly told the agent to serve the decision in front of them. Inference guesses a preference, memory honours a stated constraint.
It depends on jurisdiction and data type, so confirm specifics with privacy counsel. Commercially, the pattern that holds up is asking in the flow when the value is obvious, tying recall to a logged-in identity, and letting shoppers see, correct and clear what is stored.
Gifting, because the buyer is not the wearer and every constraint is an external fact that must survive. Alhena's 2026 study found the widest gap between two agents in jewellery, at 2.63 against 1.25, and the difference was largely whether constraints persisted.
Memory separated the leaders from the field
See exactly which memory tiers each deployment cleared, across three tiers and eleven verticals.
