Memory

DeepMind framework ↗
Human 87.9 · AI 66.5Humans lead +21.42 benchmarks →

At a glance

Holding on to information over time and retrieving it when needed.

Memory is holding on to information over time and retrieving it when needed, from general facts to what happened a few minutes ago.

The interesting comparison is across time: within a single conversation a model holds everything in context, but what it learns there is gone in the next one unless it is given notes.

Human ref.
87.9
Best AI
66.5
Gap
+21.4
Our items
11

Sub-faculties

The parts of this faculty, what our items cover today, and what is still to come

Sub-facultyItemsStatus
Delayed recall
Recalling what you studied after doing other things in between.
Mirrors Hendrycks et al. (2025)
5
in the 5-min test
Covered
Working memory
Holding a few things in mind while updating or rearranging them.
Mirrors Huang et al. (2025)
6Covered
Long-term recall
Recalling something days later.
0Not covered yet

Human vs AI

Best published result on each side, 0–100

Human ref.
87.9
Best AI
66.5

Humans. People keep track of facts scattered across months of conversation, including when things happened.

AI. Within a conversation a model holds everything in context, but memory-augmented systems still trail people over long conversations. Without external notes, what a model learns in one session is gone in the next.

Human ref. 87.9
F1 score of people answering LoCoMo's questions about a long, multi-session conversation: roughly, how much of the right answer their replies contain, from 0 to 100 (Maharana et al., 2024). About the test →
Best AI 66.5
The same questions and score for Claude Sonnet with an external memory system that stores the conversation and searches it (Tiger Data, September 2026). About the test →
255075100Perception+13Generation-35Attention+44Learning0Memory+21Reasoning-27Metacognition-55Executivefunctions+17Problemsolving+2Socialcognition+4
Human referenceOur estimateBest published AILabel = human − AI, points

Example items

Specimens written for this page. Live items are similar but never published.

Reverse the digitsspecimen · not scored

Digits appear one at a time. When they stop, type them in reverse order, last one first.

(a) Shown one at a time (0.9 s each)

94627then typed back, last one first
Figure 1. A backward digit span. Each digit shows for under a second, then disappears, so the sequence has to be held in mind. Benchmark: Huang et al. (2025). Answer: 72649: the five digits, last one first.
Huang et al. (2025) benchmark page →
Remember the wordsspecimen · not scored

Study these eight words for 20 seconds. You will be asked for them after all the other tests.

(a) Study list (20 s)

harbourpencilfalconblanketmirrorcarrotladdercomet
Figure 2. A delayed free-recall item. Recall comes several minutes and six unrelated tasks later. Benchmark: Hendrycks et al. (2025). Answer: Five or more of the eight words to pass. Extra words are logged as intrusions.
Hendrycks et al. (2025) benchmark page →

Benchmarks

2 rows
BenchmarkOriginItemsBest AIHumans
Huang et al. (2025)Huang, Sun, Wang & Dredze (2025), arXiv preprint, 2025626%—
Hendrycks et al. (2025)Hendrycks et al., A definition of AGI, 202550%—

In everyday life

  • Remembering where you parked
  • Recalling what was agreed in last week's meeting
  • Keeping a shopping list in your head