Cognitive vocabulary is the wrong frame for AI agent memory.

The question

Should AI agent memory be inspired by human memory?

The intuitive answer is yes. The problem with that answer is that these models are good enough at mimicking us that we keep reaching for human-cognitive frameworks to explain everything else they do. The vocabulary is familiar. The substrate is not.

Borrowed vocabulary

A lot of current AI agent memory concepts are imported from human cognitive psychology. The vocabulary feels right because most of us do not yet have native intuitions for how a transformer remembers, so we reach for the framework we already know. The trouble is that those words come with design intuitions attached, and those intuitions were shaped by a substrate that operates nothing like the one you are actually building on.

Three differences are worth getting straight.

Forgetting is positional, not adaptive

Humans forget what is not reinforced over time. The mechanism is biological and continuous. Use a fact, the trace strengthens; do not use it, the trace decays.

Transformers forget what sits in the middle of the context window, regardless of whether it matters. The middle decays first because of how attention distributes, not because the content there was unimportant. Stuffing more in does not add memory. It dilutes attention.

No consolidation between sessions

When humans sleep, the brain replays and abstracts the day's experiences into schemas that survive the next morning. Consolidation is a continuous biological process; the substrate keeps updating between events.

Transformers do not change between sessions. What frameworks call "consolidation" for agents is really scheduled summarisation. Useful, but not what the word implies. Treating one as the other leads to designs that expect compression to be lossless and abstraction to happen on its own. It will not.

Recall is attention, not spreading activation

Humans remember through associative cues, including things that simply happened at the same time. Two events that co-occurred can pull each other up regardless of semantic distance.

A vector store retrieves what is semantically similar to a query. Two events that happened together but read differently to an embedding model will never co-surface. That is a structural blind spot, not a tuning problem. You can pick a better embedding model; you cannot pick a model that turns proximity-in-time into proximity-in-space without explicitly adding that signal.

The real dimensions

The right design question is not what kind of memory does my agent have. It is where does my agent's attention concentrate, and what falls outside it. A different problem with a different vocabulary.

Three dimensions matter, none borrowed from cognitive science.

The teams that move fastest from here will be the ones who put the cognitive-science vocabulary down and start designing for the substrate they are actually working on.