Why AI Forgets — and What Stanford and Google Showed
AI forgets because most language models store knowledge as weights, not as records. Updating those weights to learn something new can overwrite what came before. Researchers documented this vulnerability as "catastrophic interference" in 1989 (McCloskey & Cohen, 1989). Researchers from Stanford and Google eventually showed a practical fix: give an AI agent a persistent memory stream it can retrieve from, rather than relying only on what is baked into the model (Park et al., 2023). This does not mean forgetting is solved everywhere. It means the tools to prevent it in customer-facing AI already exist and are documented.
Why does AI forget in the first place?
Most neural networks encode what they "know" as numerical weights spread across the network, not as separate, labeled facts. McCloskey and Cohen showed that training a network on new information adjusts the exact weights holding old information. Previously learned patterns are not gradually degraded. They are catastrophically lost (McCloskey & Cohen, 1989). This represents a structural reality of how the model stores knowledge, not a bug specific to any one product. Any AI attempting to "learn" a customer relationship purely by continuing to train on new conversations inherits this exact risk.
What did Stanford and Google actually show?
In a 2023 study published through Stanford and Google Research, researchers built generative agents acting as AI characters in a simulated environment. These agents kept a timestamped log of everything they observed. They retrieved the most relevant memories when deciding what to do next and periodically synthesized those memories into higher-level reflections (Park et al., 2023). The core finding was specific. Agents using this retrieve-and-reflect architecture behaved with more coherence and continuity over simulated days than agents without it. That is a real, replicable result about memory architecture in generative agents. It is not proof that any single deployed product "doesn't forget," and it is not a general theory of AI memory. It provides evidence for one workable design pattern, later reinforced by separate research on retrieval mechanics (Lewis et al., 2020).
Does retrieval alone guarantee an AI won't forget what matters?
No. A 2025 Frontiers in Psychology study examined retrieval mechanisms in generative agents directly. The researchers found that how memories are ranked and retrieved matters as much as whether they are stored at all. Poor retrieval can pull irrelevant memories forward and hurt an agent's response quality, even with an accurate memory store underneath (Hong & He, 2025). This is why "we store your conversation history" is a much weaker claim than it sounds. Storage without disciplined retrieval produces an AI that either forgets what mattered or surfaces the wrong memory at the wrong moment. Surfacing the wrong context is arguably worse than forgetting cleanly.
What should a business actually take from this research?
| Research finding | What it proves | What it does not prove |
|---|---|---|
| McCloskey & Cohen, 1989 | Networks trained sequentially can catastrophically overwrite old knowledge | That every AI system forgets this way today |
| Park et al., 2023 (Stanford/Google) | A retrieve-and-reflect memory architecture improves agent continuity | That any specific commercial product has solved memory |
| Hong & He, 2025 | Retrieval quality determines whether stored memory helps or hurts | That storing more data automatically improves outcomes |
An honest reading of this research shows forgetting in AI is an architecture problem with documented, partial solutions. It is neither fully solved nor permanently unsolvable. For call- and appointment-driven businesses, the practical question isn't "does the AI have memory" but "can it retrieve the right memory, ranked correctly, at the moment a customer calls back."
The implication
Treat "our AI remembers" as a claim to interrogate rather than a feature to take on faith. Ask exactly what is stored, how the system retrieves it, and how ranking is handled. The research proves all three matter independently. For a practical look at how memory claims get tested during a follow-up interaction, read the second phone call test. To see how persistent, retrievable memory operates inside a live product, visit AiPRL OS.
Keep reading: our view of omnichannel.