Consolidation

Example

You work through something difficult in the evening, sleep, and find the next morning that you handle it better, without having practised again. Something happened to the memory while you weren't using it.

Consolidation in Humans

Definition

The transformation of a memory over the hours and days after acquisition, through recurrent reactivation during both waking and sleep (⁠Dudai et al., 2015⁠). The trace does not simply settle. It is redistributed to additional sites and integrated into existing knowledge, so what is available later differs from what was encoded.

*Framing note: scoped to the post-acquisition transformation of a specific memory. Schema Formation covers the resulting knowledge structures; Reconsolidation covers re-entry into a labile state after retrieval. The three are aspects of one process rather than a fixed sequence. Reconsolidation is conditional: it does not appear to follow every reactivation of a long-term memory, and is more likely when new information is available in the retrieval situation and when the reactivated representation is strong enough to control behaviour (⁠Dudai et al., 2015⁠). They are separated here because each stands in a different relation to the AI material.*

Neural Substrate

Human evidence exists at the systems level and is the basis for what follows. Intracranial recordings from patients with epilepsy identified recurring spatiotemporal sequences of high gamma activity spanning widespread cortical sites during waking, and in all six patients analysed those sequences recurred more often across the two nights following the waking period than across the two preceding it; more than 90% of the recurring sequences were novel, and their recurrence was coupled to sleep spindles, down-to-up transitions, theta bursts, and hippocampal sharp wave ripples (⁠Jiang et al., 2017⁠). Scalp EEG in 20 healthy sleepers shows the same reactivation without any experimenter-delivered cue, with classifier evidence for the stimulus category studied before a nap rising above chance between 800 and 1200 ms after the slow oscillation downstate, and only where a slow oscillation and a spindle occurred together, since neither event alone produced above-chance classification (⁠Schreiner et al., 2021⁠). Human fMRI across three months shows the redistribution itself, with hippocampal activity during confident recognition of remote items declining while ventral medial prefrontal activity rises, the two closely anticorrelated, and with the amount of slow-wave sleep during a post-encoding nap predicting retention for material learned before the nap but not after it (⁠Takashima et al., 2006⁠). Neither the intracranial nor the scalp EEG study establishes that reactivation causes consolidation; Jiang et al. say so explicitly, and Schreiner et al. report a correlation between reactivation strength and retention.

Synaptic consolidation is a different case, because no human evidence characterises it. The account of it in the reviews is drawn from invertebrate and rodent preparations, chiefly Aplysia and mice, with the classic time course established in goldfish (⁠Dudai, 2004⁠; ⁠Dudai et al., 2015⁠). At that level the claim is not regional: consolidation is a molecular and cellular process at the circuits where the memory was first encoded, traditionally assumed to draw to a close within hours of learning, and it has been described in every species, preparation and task examined so far. Systems consolidation is the slower time-dependent reorganisation of the representation across distributed circuits, over weeks to years, during which the trace spreads to new locations while relinquishing dependence on parts of the circuit that supported its acquisition (⁠Dudai, 2004⁠). Dudai et al. (2015) treat synaptic consolidation as a repeating subroutine within the systems-level process rather than as a separate mechanism, which is why the non-human basis of the first does not carry over to the second.

AI at a Glance

In humans, consolidation transforms a memory over hours to years through recurrent reactivation during waking and sleep, redistributing it across brain sites and integrating it into existing knowledge (⁠Dudai et al., 2015⁠; ⁠Takashima et al., 2006⁠). In AI systems an offline pass over already-stored material is implemented and measured at the memory layer, either as prompted reorganisation of a retained context before queries arrive (⁠Lin et al., 2025⁠) or as scheduled deduplication and promotion inside a pipeline whose stages are named for the human process (⁠Kerestecioglu et al., 2026⁠), but the selection principle differs, since what gets worked on is chosen by predicted usefulness to an anticipated question rather than by significance or by fit to existing knowledge.

Context: the body of text a system is given alongside a question, such as a document, a codebase, or a conversation history (⁠Lin et al., 2025⁠).
Memory layer: the component of a deployed system that stores material outside the model itself and decides what to retain, rewrite, or retrieve, as distinct from the model's trained parameters (⁠Kerestecioglu et al., 2026⁠).

Consolidation in AI

How Far It Has Been Measured

Measurement in AI: Indirectly Measured

AI Memory Performance: Not Yet Compared

Where It Lives in the System

What Has Been Measured

Widely used AI assistants update nothing in the model itself between conversations; a 2026 preprint tested nightly consolidation of conversations into the model's weights (its long-term learned parameters) and found it retained 80.4% of knowledge versus 36.8% for repeatedly summarizing the conversation, across 10 conversations and 1,146 questions (⁠Dennis et al., 2026⁠).
Other 2026 preprints build explicit 'sleep' phases into research models, distilling recent in-context knowledge into long-term parameters with replay, or converting the recent conversation into persistent weights offline, and report better performance with longer sleep; these are prototypes scored against other AI methods, not tests against people (Behrouz et al., 2026; Lee et al., 2026).
In training, facts a model picks up are not stabilized: each exposure raises their probability, but later training dilutes it through forgetting that follows a power law, measured as knowledge acquisition and forgetting rather than as consolidation (⁠Chang et al., 2024⁠).
In saved memory, systems such as Mem0 extract and 'consolidate' facts from conversations into a store the AI can search later; this is scored only as part of broad long-conversation question-answering benchmarks, where Mem0 beat OpenAI's memory feature by 26% (relative) on a model-judged score (⁠Chhikara et al., 2025⁠).
A 2026 preprint found that rewriting interactions into stored facts kept claims but erased the limits their source placed on using them (for example, a user's say-so becoming a standing instruction) in 48 of 49 combinations of seven memory systems and seven language models, so the failure was nearly uniform across systems; this resembles consolidation losing source detail but included no human comparison (⁠Zhan et al., 2026⁠).
The closest classic paradigm tried on AI is interference: across 39 language models, earlier information in the conversation disrupted recall of later updates more than the reverse (d = 1.73), opposite to the typical human pattern, but this tests the context window (what the model can see in the current conversation) with no delay or offline period, and no people were tested (⁠Chattaraj & Raj, 2026⁠).
Not yet compared: no study was found that gave AI and people the same consolidation test, such as retention after an offline period versus none; the AI results are gains of one engineered method over another, which cannot be placed on a common scale with the human sleep benefit of roughly g = 0.3 to 0.4 (⁠Berres & Erdfelder, 2021⁠).