AI assistants can lose or fail to apply information over sustained interactions, but the claim that they ignore you 90% of the time is not supported by the available evidence. Long-term memory benchmarks measure specific abilities—such as recalling facts, updating knowledge, and using preferences during a task—and show that remembering something and acting on it are different challenges. A shared memory could help with continuity, but no source here establishes that one memory fixes most failures across assistants.
Is it true that AI ignores what you tell it 90% of the time?
No general 90% failure rate is established by the available studies. That figure should not be treated as a measured rate for AI assistants overall. The relevant research tests defined tasks and systems; it does not measure every instruction a person gives an assistant in everyday use.
As an Amazon Associate I earn from qualifying purchases.
LongMemEval, an ICLR benchmark paper published in 2025, reports a 30% accuracy drop for commercial chat assistants and long-context language models on memorizing information across sustained interactions. That result describes performance on the benchmark’s memorization task, not the percentage of user instructions ignored in general.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →What does “remembering” mean for an AI assistant?
Long-term memory is not a single skill. A system might retrieve a fact when directly asked but fail to use it when choosing an action. It might also rely on an old detail after circumstances have changed. LongMemEval organizes evaluation around five abilities:
- Information extraction: identifying details in prior interactions.
- Multi-session reasoning: combining information from separate conversations.
- Temporal reasoning: understanding when a detail applied or changed.
- Knowledge updates: incorporating newer information rather than relying on an outdated version.
- Abstention: recognizing when the available memory is insufficient to answer.
These distinctions help explain why “I told it before” does not guarantee a useful response later: retrieving a detail, deciding it is relevant, and applying it correctly are separate steps.
Why can an assistant recall a preference but still fail to use it?
Recall is retrospective: the system is asked about something that happened earlier. Many real tasks require prospective memory: the assistant must notice that an earlier constraint matters now and act on it without being reminded. Microsoft Research’s TriggerBench, published in June 2026, studies this ability to spontaneously recall and act on latent constraints; its abstract reports that prospective memory is more difficult than retrospective memory in its evaluation.
Rank #2
Mem2ActBench examines another part of the gap: active use of memory in tool-oriented tasks, including selecting tools and grounding their parameters in relevant information. The authors report that current systems remain inadequate at this kind of memory use. A remembered preference is therefore not the same as a preference reliably shaping a later action.
Recommended Free Tools
What recent benchmarks actually test
| Benchmark | Focus | Reported evidence |
|---|---|---|
| LongMemEval, ICLR, 2025 | Five long-term memory abilities: extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention. | Reports a 30% accuracy drop on memorizing information across sustained interactions for commercial chat assistants and long-context LLMs. |
| Mem2ActBench, ACL Anthology, 2026 | Actively applying memory to tool use, including tool choice and parameter grounding. | Authors describe 2,029 synthesized sessions and 400 tool-use tasks. Human evaluation judged 91.3% of generated tasks strongly memory-dependent. Experiments on seven memory frameworks found current systems inadequate at active memory use. |
| PersistBench, PMLR, 2026 | Risks from long-term memory, including inappropriate carryover between domains and reinforcement of user biases. | Across 18 evaluated models, the authors report median failure rates of 53% on cross-domain samples and 97% on memory-induced sycophancy samples. These are benchmark-sample results, not everyday-use rates. |
| TriggerBench, Microsoft Research, June 2026 | Prospective memory: spontaneously recalling and acting on latent constraints. | The abstract says prospective memory is more difficult than retrospective memory in the evaluation; no numeric result is stated here. |
| RHELM, Microsoft Research | Realistic, heterogeneous, evolving long-horizon assistant memory. | The benchmark website describes its evaluation focus; no performance statistic is stated here. |
Can one shared memory work across different AI assistants?
It is a plausible design idea, not a demonstrated fix in the evidence above. The cited benchmarks examine memory abilities, task execution, or risks; they do not establish that a single memory shared across assistants resolves most instruction failures. Nor do they verify which named products currently work across assistants or what controls those products provide.
If you are assessing a shared-memory tool or feature, treat these as questions to verify rather than assumed capabilities:
- Coverage: Which assistants, models, and workflows can actually read and use the memory?
- Control: Can you inspect, edit, delete, or correct saved details?
- Relevance and forgetting: Can outdated, uncertain, or unrelated information be withheld or removed?
- Application: Does the system merely retrieve facts, or does it use preferences correctly during tasks?
- Data handling: Where is information stored, and how is it used? Check the specific provider’s current policy rather than assuming.
These are practical evaluation criteria, not product endorsements or features verified for any particular tool.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to get more reliable results from an assistant
Until you know how a particular assistant handles memory, make important constraints visible at the point where they matter. These steps are practical safeguards, not guarantees that an assistant will retain or apply information:
Free tools Windows power users keep installed
One-click scans. No signup required.
- Restate high-impact constraints in the active request. Include requirements that would materially change the answer, rather than relying on an earlier conversation.
- Separate durable preferences from one-time instructions. Say whether a preference should apply generally or only to the current task.
- Make updates explicit. If your circumstances or preferences change, state what is new and what it replaces.
- Check consequential actions before they happen. Ask the assistant to show the assumptions or parameters it is using before sending, booking, editing, or otherwise acting.
- Correct carryover that does not belong. If the assistant brings an unrelated or outdated detail into a new topic, tell it that the detail is irrelevant or no longer current, and use any available memory controls to correct or remove it.
Why more memory is not always better
Persistent memory can improve continuity, but it can also carry information into the wrong context or reinforce a user’s existing bias. PersistBench names these risks cross-domain leakage—inappropriately injecting stored context into another domain—and memory-induced sycophancy—using stored memories in ways that reinforce user biases. Its benchmark results are warnings about failure modes to evaluate, not estimates of how often ordinary users will encounter them.
A useful memory system therefore needs more than a large store of past details. It must also decide what is relevant, account for updates, and avoid using a remembered fact simply because it is available.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
