Story codec — scene · voice · word¶

A story compresses to three redundant anchors — SCENE (visual+spatial), VOICE (auditory+character), WORD (semantic+lexical). Any one leg recovers the others, because human memory is associative. Thirty words can index a thousand-page story for the right reader.

🌱 seedling tended 2026-05-09 compression memory mnemonic codec notation

flowchart LR
  scene[SCENE] --> story((the story))
  voice[VOICE] --> story
  word[WORD]   --> story
  scene -.recovers.-> voice
  voice -.recovers.-> word
  word  -.recovers.-> scene

L0 — TL;DR (≤5 lines)¶

A story compresses to three anchors — SCENE (visual + spatial + sensory), VOICE (auditory + character), WORD (semantic + lexical). Each leg independently indexes the story; remember any one and the network reconstructs the rest, because human memory is associative across modalities. Thirty written words can hold a thousand-page novel for someone who has read it. The catalog below is the format, the rules for composing one, and the failure modes.

L1 — Overview¶

Core question¶

What's the smallest, most recoverable representation of an arbitrary story (a memory, a book, a relationship, a repo state, a dream) such that any single anchor reconstructs it, and the format works for both humans and machines?

Why it matters¶

Notes are read sequentially; cues retrieve associatively.
Most lived stories aren't lost — their retrieval cues are.
A redundant cue is a backup for itself: if the smell fades, the voice is still there; if the voice fades, the word is still there.
The repo's whole language is compression-with-purpose (see COMPRESSIONS.md). This is the entry that handles narratives — what proverbs handle for decisions.

Mermaid map (L1)¶

flowchart LR
  experience[lived/read/imagined story] --> encode[encode triple]
  encode --> scene[SCENE · place · object · motion · smell]
  encode --> voice[VOICE · pace · register · tic]
  encode --> word[WORD · the resonant token]
  scene --> recall[any leg → story]
  voice --> recall
  word  --> recall
  recall --> reconstruct[narrative re-emerges with detail]

The triple at a glance¶

Leg	Modality engaged	Typical length	What it carries
SCENE	visual + spatial + sensory (smell, touch, temperature)	5–15 words	one place, one object, one motion, one sense
VOICE	auditory + character	5–10 words	pace, register, tic — how someone said something
WORD	semantic + lexical	1–3 words	the resonant token (a name, a phrase, the title)

Three legs is not arbitrary. The brain has at least three orthogonal imagery channels that stack multiplicatively (see investigations/BODY-AS-ENGINE.md#7-brain-imagery-channels): visuospatial sketchpad, phonological loop, and semantic/lexical memory. The triple is one anchor in each.

L2 — Deep dive¶

The notation¶

Inline form (for prose, comments, frontier cards):

~ SCENE | VOICE > WORD

Concretely:

~ harbor at dusk · gulls · salt · fisherman missing a finger | gravelly · slow · ends sentences early > "almost"

Block form (for memory files, glossary, lesson capsules):

@cue
  scene: harbor at dusk · gulls · salt · fisherman missing a finger
  voice: gravelly · slow · ends sentences early
  word:  "almost"
  story: stories/2024-summer-portsmouth.md   # optional pointer
  cap:   1300pp → 35w                        # optional ratio

Single-line strict form (machine-parseable):

[scene: ...] [voice: ...] [word: ...]

Punctuation rules:

~ opens the scene, | enters the voice, > lands the word.
· separates tokens within a leg. (Middle-dot / · — same one used by the Mermaid L1 cards.)
Quotes around the WORD when it is itself a phrase a character said.
Em-dashes inside SCENE are fine; periods aren't (a leg is one breath).

How to compose one¶

Five steps, in this order:

SCENE first. Pick a single physical location. Add one object the hand could touch, one motion (a person crossing, a wave, a cup being set down), and one non-visual sense (smell of pine, cold metal railing, the heat from a stove). Avoid abstractions ("happiness", "the meeting") — they don't anchor.
VOICE second. Pick one speaker — a person, the wind, an engine, the narrator's own internal monologue — and capture how they sounded with three tokens: pace (slow / clipped / urgent), register (warm / formal / brittle), and a tic (a verbal habit, a pause, a way of starting sentences). The tic is the recovery handle.
WORD third. The single resonant token. It is the name only the story uses, the phrase that recurs, the punchline, the title. If it's also a common word ("yes", "OK"), the cue won't retrieve.
Stress test. Hide two legs. Show one to someone who knows the story and ask them to reconstruct. If they can't, the leg is too thin or too generic — fix and retry.
Tend. Re-read the triple a week later. If it still pulls the story, it's tended. If it doesn't, the encoding was decorative rather than diagnostic — recompose.

Why three legs (and not two, or one)¶

Redundancy — any leg recovers the others. This is RAID-1 for human memory: no single cue is load-bearing. Lose the smell, the word still works. Forget the word, the voice will pull it back.
Multimodal stacking — the three channels (visual+spatial, auditory+character, semantic+lexical) are processed by different neural circuits and compete minimally. Two legs in the same channel don't stack — they cancel (see BODY-AS-ENGINE.md#8-stacking).
Composition — scene gives "where", voice gives "who", word gives "what". The narrative skeleton emerges naturally from the triple without naming events explicitly.
One leg might be culturally absent. Aphantasic readers (~2–5 % of people) can't form vivid scenes; the VOICE and WORD legs carry the cue for them. People with weak inner speech rely on SCENE and VOICE. Three is the minimum that survives one channel-failure.

Worked examples¶

Triple	What it cues
`~ harbor · dusk · gulls · cold rope under the hand · fisherman missing a finger \\| gravelly · slow · ends sentences early > "almost"`	a remembered week somewhere, with one specific person
`~ wide ballroom · candle smoke · dust in beams · a girl seeing a man across the room \\| stammered · earnest · over-apologizing > "comet"`	War and Peace, for someone who's read it
`~ kitchen 2am · linoleum · a kettle whistling · a phone face-down on the counter \\| flat · medical · pauses before names > "results"`	a specific bad night, from one's own life
`~ open-plan office · whiteboard · a Slack ping · the laptop fan getting louder \\| brisk · over-confident · says "actually" first > "P0"`	a particular incident at a job
`~ forest path · mid-morning · pine resin · a child running ahead with a stick \\| breathless · giggling · ends every line in "..." > "look!"`	a memory you have but rarely retrieve
`~ commit log · 03:17 UTC · pre-commit hook screaming · git diff scrolling \\| dry · self-deprecating · says "well that escalated" > "F118"`	a repo state — a session in this very project
`~ dream beach · purple sky · sand that walks · a clock with no hands \\| your own voice but lower · slow · loops back > "again"`	a recurring dream

Failure modes¶

Failure	What it looks like	Fix
Abstract scene	`~ a feeling of dread · regret · the past`	replace with one room, one object, one smell
Generic voice	`~ ... \\| nice guy, friendly`	add a tic — "starts every answer with a long inhale"
Empty word	`~ ... > "yes"`	swap to a name, an unusual phrase, or the title
All three same channel	`~ paragraph 1 \\| paragraph 2 > paragraph 3` (all semantic)	split across modalities
Decorative triple	sounds nice but pulls nothing back	recompose by re-entering the memory and noticing what first arrives
No underlying story	the triple is a poem, not a cue	acknowledge — store as a poem
Cue used for two stories	one triple retrieving competing memories	disambiguate one leg (a unique smell, a unique word)
Tic that has aged out	the speaker has changed; the voice no longer matches	re-encode after the change

How to test a triple¶

Three tests, in increasing strength:

Self-test (one week later): read only the triple. Did the story re-emerge with detail beyond the triple itself? If yes, the cue is live.
Friend-test: hide two legs from someone who lived/read the same story. Show them one. Ask what comes back. The good triples produce "oh — that time." The bad ones produce a shrug.
Time-decay test (one year later): re-read. If only one leg still pulls the story, the redundancy did its job — but recompose. If none do, the story has either consolidated to long-term (good) or faded (record now what's left).

Where it fits in the godding codec stack¶

COMPRESSIONS.md is the catalog of every codec the repo uses. The cue triple is a new row.

Codec	Compresses	Ratio (rough)	Where
TL;DR / L0	a page	100 : 1	every page top
Mermaid L1	structure	5 : 1	overview
Glossary	a concept	a definition + tooltip	`GLOSSARY.md`
Fun-fact metaphor	a hard concept	held-object metaphor	`FUN-FACTS.md`
Proverb	a decision rule	5–15 words	`PROVERBS.md`
Cue triple	a story	~30 words	lessons · frontier cards · personal logs

Each is a different lossy codec for a different kind of source. Stories don't fit proverbs (a proverb is a rule, not an episode); diagrams don't fit cue triples (a flow has structure, not narrative voice). The codecs are orthogonal — pick by source type.

Machine form¶

For repo-wide tooling, the strict YAML form is the canonical encoding:

- cue:
    scene: "harbor at dusk · gulls · salt · fisherman missing a finger"
    voice: "gravelly · slow · ends sentences early"
    word:  "almost"
  refs:
    story:  "logs/2024-08-10.md"
    lesson: "memory/LESSONS.md#L-742"
  meta:
    cap:    "1300pp → 35w"
    tended: "2026-05-09"

A small parser (regex-light) can:

Extract triples from prose for indexing.
Verify each leg's length contract (5–15 / 5–10 / 1–3 words).
Detect failure modes (all-abstract, all-same-channel, generic word).
Build a reverse index: word → triple → story-ref.

For models specifically: a triple is a high-entropy retrieval prompt. "Recall the story cued by ~ harbor at dusk · gulls · salt · ... \| gravelly · slow · ... > 'almost'" lands cleaner than free prose because the three legs constrain the search space sharply.

Where to use it in this repo¶

Lesson capsules (memory/INDEX.md) — attach a triple to lessons that are episodic ("the time the staging pipeline burned down") rather than rule-shaped. Individual lesson files live under memory/lessons/L-NNN.md (not mirrored to the site — browse the repo for the full set).
Frontier cards (tasks/FRONTIER.md) — triples that cue the origin of an open question.
Decision records — a triple anchors why a decision was made, not just what it was.
Personal logs / runlog notes — the triple replaces 200 words of context-restoration when you reopen old work.
Onboarding stories — a few triples that cue the canonical failures of the project; faster than reading the full lesson list.

Design notes (for future variants)¶

Two-leg cues (e.g. SCENE + WORD) work for short stories where the speaker doesn't matter. The redundancy drops to RAID-0 — one failure loses the cue. Fine for low-stakes recall.
Four-leg variants (adding KINESTHETIC: a body sensation — warmth in the chest, knot in the stomach) are worth experimenting with for very emotionally charged stories. Cost: harder to compose, longer to read.
Group triples — a shared triple for a team's memory of a project. The voice leg shifts from one speaker to "the room", the word leg becomes the team's coined term. Tested informally; works.

Open questions¶

Empirical recall fidelity vs. triple length. Does a 30-word triple retrieve more than a 60-word free summary? (Anecdotal yes; no controlled test in this repo yet.)
Cross-cultural transfer. The triple format assumes a literate, internal-monologue reader. Songlines and oral epics use longer, rhythmic structures — is the triple a Western shortcut?
Aphantasic compensation patterns. Which two-leg fallbacks (VOICE+WORD, SCENE+WORD) hold up at one-year decay?
Machine-cue performance. Do LLMs reconstruct stories from triples more faithfully than from free summaries of the same length?

References¶

Yates, F. A. (1966). The Art of Memory — method of loci as the oldest known scene-anchor codec.
Bruce Chatwin (1987). The Songlines — Aboriginal cue-and-recover geography as an existence proof at continental scale.
Baddeley, A. (2003). Working memory: looking back and looking forward — phonological loop / visuospatial sketchpad split.
Cowan, N. (2010). The magical mystery four — working-memory capacity revision.
Proust, Swann's Way — olfactory cue (madeleine) recovering an hour of childhood, in literary form.
Tulving, E. (1972). Episodic and semantic memory — the encoding vocabulary the triple sits inside.
Daniel Kahneman, Thinking, Fast and Slow — System 1 retrieval is cue-driven, not search-driven.