Your daughter texts a photo: a steak the size of a small continent, a mountain of mashed potato, a glass of milk. Anyone else sees a plate. You see a private joke that has been running since she was little — the meal you both treat as honest food, the line you used to say when vegetables were propaganda and protein was the plot.

You type back: healthy but yummy.

Three words. Not a caption. Not a summary. Not an AI essay about fatherhood. A relational cue — a compression code that only works inside one relationship, and that, if it ever comes back to you at ninety when the photo arrives again, is the only answer worth retrieving.

That is the product. Everything else is plumbing.

The wrong product category

Most “AI memory” demos still sell the wrong unit. They store more chat. They summarise your week. They caption your photos. They answer questions about you in full sentences, as if the goal were a second biographer living in your pocket.

Healthy aging — and ordinary midlife busyness — do not primarily destroy storage. They thin out self-initiated retrieval. Free recall gets harder; recognition, when the right support is present, holds up better.^1,2 Tip-of-the-tongue moments often look like missing files and behave like weak links: supply a good cue and the target returns.³

So the job is not “remember for me.” The job is: hand me a chink while I am still inside the thought that needs it, then shut up so the flood can be mine.

What happened with rsync

I was in the middle of ordinary work — moving data with rsync — when a reconstruction landed in conversational time: Samba, Andrew “Tridge” Tridgell, the tiny Australian computer-chess scene, Joel Veness down the road in Sydney, my own engine Chompster, KnightCap’s learning curve. Names I half-held snapped into a shape I did not have fully assembled in working memory.

Two things mattered more than the gossip value.

First, timing. It was not a weekend report. It arrived while the thought that summoned it was still alive — near instant, in the same cycle. That is what makes a reconstruction into augmentation rather than homework. Call it conversational time, or in-loop cognition: the result re-enters the activation window of autobiographical memory instead of arriving after the window has closed.⁵

Second, reciprocal graph completion. I held private high-semantic edges no search engine owns (“I knew that bloke,” “Joel lived down the road,” “we met”). The machine held breadth and public join power. Neither side had the full shape. The shape emerged between us.

That is not “the AI knows your life.” That is a joint graph being completed under a latency constraint.

The Recognition Loop

Memory augmentation works when the machine returns a minimal relational cue inside the activation window of the thought that summoned it. So the system has three jobs — and each fails without the other two.

Job 1 — Mine: intimacy builds compression codebooks

Families invent private dialects. Film quotes. Mispronounced childhood words. Food names. One-word references to disasters. To an outsider they are nearly information-free. Inside the relationship they reinstate years of shared state.

Count-based importance is the wrong ranking function for this layer. What you want is closer to:

low frequency
+ large temporal span
+ same relationship
+ contextual reuse
= high autobiographical significance

The rare phrase that reappears years apart inside one relationship is dear. The high-frequency “see you soon” is not.

This is the personal application of a doctrine already written for code and corpora: description is often regenerable; significance is the expensive thing to keep.⁶ For a photo, “steak, mashed potatoes and milk” is cheap. “Healthy but yummy” is not in the pixels. It lives at the intersection of the photo, two people’s history, shared language, and the present moment.

Encoding specificity is the science rhyme: what was stored determines which cues can reopen it.⁷ Caption the exhibit and you store one thing. Mine the relationship codebook and you store another.

Job 2 — Surface almost nothing: the edge surfacer

The mature interface is not a memory chatbot. It is an edge surfacer.

You are in a cognitive neighbourhood — rsync on the terminal, a daughter’s photo on the phone, a name half-formed. The system watches that neighbourhood against a compiled wiki of your life and occasionally places one high-potential edge string where you can see it:

Then it stops.

No: “Scott, did you know that Andrew Tridgell was involved in…”

You know. You need the chink. Recognition does the rest — the same interaction contract that memory books and reminiscence practice have used by hand for years: prompts and artefacts, not a machine that performs your autobiography for you.^8,9

Rank the anti-patterns by how hard they violate the dignity constraint:

1. Narrator — tells the story of the meal in full sentences.
2. Interpreter — explains what the meal means about your parenting.
3. Oracle — asserts your feelings and history as settled fact over your head.

The dignity rule is architectural, not tone: store relationships and pointers, not verdicts; show the exhibit or the cue, not the conclusion; offer candidates, then yield.⁴ The machine hands you a key to a room that belongs to you. It does not walk through the room narrating the furniture.

Job 3 — Time: compile offline, place online

Deep graph walks at the moment of need are the wrong architecture for autobiographical windows. The thought that needs a cue is short-lived. The archive that contains the cue is large.

So: spend cognition offline. Homogenise email, messages, photos-as-semantic-twins, public traces, project notes into a personal wiki. Let a janitor process leave significance: unresolved when the system cannot yet know why a photo matters — honesty beats a fake caption. At live time, do almost no work: match active context to a pre-mined cue and place it before the window closes.

Capture was never the bottleneck. Pendants that only record from the day you clip them on miss the prehistory. The exhaust is already on your devices — decades of email, photos, messages. Compilation is the missing step.⁴

The photo twin

Do not put the JPEG in the wiki. Put its semantic twin:

## Observed
- large steak, mashed potato, glass of milk

## People
- [[Mackenzie]]

## Significance
- unresolved   # until another corpus supplies the edge

## Cue
- (empty until mined)

## Edges
- [[Mackenzie]] · family meals · childhood food language

## Sources
- messages://…

Later, the conversation where you say “healthy but yummy,” plus older history, can fill significance and cue. That is synthetic augmentation without falsifying the raw exhibit: the photo stays the photo; the compiled layer holds meaning and the way back.

What success looks like at ninety

Mackenzie sends the plate again. A dumb assistant says: “Mackenzie sent a photo of steak, mashed potatoes and milk.” A slightly smarter one invents a nutrition lecture. A dangerous oracle tells you what you feel about your daughter.

MemoryMate says:

And stops.

If the flood comes — the little voice, the teaching, the private food philosophy — the system worked. If you need a second candidate cue, offer another chink. Never escalate into recitation because silence felt underwhelming to the product manager.

Design the loop, not the chatbot

If you are building this — for yourself, a parent, a product — design the three stages as one system:

• Relationship compiler — mine temporal recurrence over one relationship at a time; keep the significance formula honest; allow unresolved.
• Edge surfacer — peripheral, minimal, yield-first; rank and ban narrator/interpreter/oracle modes for live autobiography.
• In-loop latency budget — measure success as “cue arrived while the thought was still active,” not as “answer was thorough.”

The Life Wiki established cues-not-conclusions at query time.⁴ Recognition applied to a career narrative is a sibling move.¹⁰ This piece adds the rest of the loop: proactive ambient edges, the relational cue as an object, and timing as a first-class job — still scoped to the personal past, still leaving the remembering to the person who lived it.

You do not need a whole sentence. You need the edges. And you need them while you are still the kind of animal that can recognise yourself in them.