Skip to content

Brain memory management

Working memory is small (~3-7 slots). Long-term is large but cue-only. Sleep is the consolidation routine that prunes and re-files what you took in.
🪨 stub tended 2026-05-07 research memory sleep working-memory 
flowchart LR
  input[day input] --> wm[(working memory: 3-7 slots)]
  wm --> sl[sleep]
  sl --> ltm[(long-term · cue-only)]
  ltm -.recall by cue.-> wm
Connected work

Investigation · rating: medium. L0/L1 below carry the working compression; L2 is empty.

Status: stub | 2026-05-07 | rating: medium Compress levels: L0 ↓ L1 ↓ L2 (empty)

L0 — TL;DR (≤5 lines)

Working memory is small (~3–7 slots). Long-term memory is large but only accessible by cue. Sleep is the consolidation/regression routine that prunes and re-files what you took in. Without sleep, the stack overflows and behaviour decoheres — "you go crazy" is not a metaphor. External traces (paintings, notes, layouts — see STIGMERGY) extend the slot count cheaply, by replacing recall with re-perception.

L1 — Overview

Core question

Given a finite working-memory stack, finite daily energy, and sleep as the only consolidation channel, what management policies actually work — and which everyday practices (writing, single-scene imagining, returning to physical objects) are good implementations of those policies?

Why it matters

  • Every "I forgot" event is a stack-management failure or a cue failure.
  • Sleep is non-negotiable but its quality is tunable.
  • Most productivity advice is implicit memory management; making it explicit lets you pick the cheapest version per energy unit.
  • This is the human analogue of the swarm's compact.py and memory/INDEX.md.

Mermaid map (L1)

flowchart LR
  in[Sensory input] --> wm[Working memory ~7 slots]
  wm -->|active rehearsal| wm
  wm -->|write| trace[External trace]
  wm -->|encode| ltm[Long-term memory]
  trace -->|cue| wm
  cue[External cue] -->|retrieve| ltm
  ltm -->|recall| wm
  sleep[Sleep] -.consolidation.-> ltm
  sleep -.pruning.-> ltm

Sleep is a side-channel that runs offline; everything else is online.

Skeleton sub-claims

  1. Slot count is the binding constraint. Treat it like a scarce resource — don't burn slots on things you can re-perceive (calendar, layout, sticky note).
  2. Cued recall ≫ free recall. Designing your environment to surface the right cues at the right time is the highest-leverage memory practice.
  3. Sleep does both compression and pruning. Consolidating signal and discarding noise are coupled; you cannot do one without the other.
  4. Writing is the cheapest reliable encode. Higher fidelity than passive intention, lower energy than full visualization.
  5. Imagining a single scene (rather than a movie) is energy-efficient cue laying — the single image acts as a future stigmergic trace inside your own head.
  6. Without sleep the system fails predictably: irritability → confusion → hallucination → physical breakdown. The order is mechanism, not folklore.

L2 — Deep dive

The slot count is the actual constraint

Working memory is often quoted as 7 ± 2 (Miller 1956), but lab measurements under realistic load put the functional count at 3–4 chunks. The number moves with: emotional state (lower under stress), familiarity (higher when chunks are well-encoded), and modality (visual vs verbal slots overlap partially, so mixing modalities can buy a slot back).

The practical implication is brutal: if you are holding three things in mind, asking yourself a fourth question often evicts one of the first three. You will not notice the eviction until you reach for it. The ubiquity of "wait, what was I doing?" is not a personal failing — it is a predictable consequence of slot pressure.

The corollary is the entire success of stigmergy (STIGMERGY-IN-DAILY-LIFE): every external trace is a slot you do not have to occupy.

Cued recall is cheap; free recall is expensive

The brain has vast storage and a tiny query interface. Free recall ("name a French river") is computationally expensive and unreliable. Cued recall ("you visited X last summer — what river was that?") is fast and accurate.

This shapes daily decisions without us realizing: - Why grocery lists work: the list is a sequence of cues, each retrieving the recipe context it was written in. - Why returning to a room fixes the "why did I come in here" problem: the original room is the cue that encoded the intention. - Why visualizing a single scene works (covered in LEARNABLE-SKILLS-FOR-VARIANCE): a static, vivid image is a high-quality cue that retrieves itself.

The actionable form: when you need to remember something, design the cue that will retrieve it, then place the cue where future-you will encounter it. Don't try to "remember harder."

Sleep is a coupled compression-and-pruning routine

Slow-wave sleep replays neural patterns that fired during the day, with some patterns reinforced (consolidation) and others weakened (pruning). The two are not separable — the same replay event that consolidates one trace weakens its less-rehearsed neighbours.

Two consequences:

  1. A sleep-deprived day's memory is unstable not because nothing was encoded, but because nothing was filed. The traces are there but not indexed. Recall fails, then improves after a single night's sleep, then slowly stabilizes.
  2. Long sleep deprivation produces hallucination, not just fatigue. The pruning step fails — patterns that should have been suppressed continue firing alongside intended ones. Behaviour decoheres in a textbook compression-system-failure pattern.

The advice "sleep on it" is a literal, mechanistic instruction.

The energy-cost of imagining a movie vs a single scene

Vivid visualisation is metabolically expensive — fMRI shows recruitment of visual cortex similar to actual perception. A narrative visualisation (running a movie of tomorrow) costs more again, because each frame must be generated and the prior frames maintained.

A single still image gets most of the benefit for a fraction of the cost. The brain compresses the rehearsed movie down to the image anyway in the consolidation phase. Front-loading the compression — choosing one vivid frame and laying it down — saves the energy and gives a better cue.

This connects directly to LEARNABLE-SKILLS-FOR-VARIANCE drill #4.

Open questions

  • What is the actual slot count under real-world load (vs lab estimates)?
  • Is single-scene imagining measurably cheaper than narrative rehearsal in EEG/fMRI?
  • Can a daily "compact" routine (10 min reflective writing) substitute for some sleep loss?
  • How does multi-modal cueing (smell, place, posture) compare to verbal cueing on retention?

References

  • (Pending L2 fill — Miller 1956 (slot count), Tononi & Cirelli (sleep & synaptic homeostasis), Walker (Why We Sleep) are obvious entry points; verify before citing.)

Inspiration sources

  • The user's framing: "sleep is like a regression routine, without sleep you go crazy." That phrasing maps the brain to a software system with a recurring batch job — useful metaphor, holds up reasonably well.
  • Wikipedia article on working memory.

See also