Skip to content
therobot.is

Live demo

Ask the memory

You supply the words and the state the question is asked in — who you are with, where and when, and the mood and hormone levels the request carries. Five retrieval paths answer it at once: meaning, language, context, affect and the association graph. Each abstains rather than guesses, their rankings are fused, and every result names the paths that surfaced it.

How inference works, on the landing →

starting the retrieval paths…

A retrieval, end to end

From inputs to a synthesis retrieval alone would not reach.

A simulated session. Task, conversation, state and an image go in. The space is searched, a shortlist is scored, and a sieve passes what the task needs while keeping a strong tangential. The agent synthesises something the direct records could not yield, then scores every emitted memory for usefulness, and the graph is reshaped by the result. The next recall is different because this one happened.

inputs

embodiment

desktop companion · voice + webcam + screen

The person the agent is talking to, seen through the session's webcam.webcam frame · derived mood below
where

agent: cloud · us-east — human: New York, NY

supplied by the session, never inferred

when

2026-09-18 · 19:40 EDT · evening

problem

week-one churn — 62% gone after the first session

with

Keith · product lead · compartment: product/team

artifact

retention section of the spec

conversation
  1. Keiththe numbers are ugly. 62% gone after the first session.
  2. AdaEmpty dashboard on first load — nothing to do.
  3. KeithWe have a checklist draft. It's not landing.
  4. AdaThe interviews say “I didn't know what to do first.”
  5. KeithI keep thinking about how games handle nothing happening.
  6. AdaSay more.
state

reflective · low arousal

v 0.35 · a 0.22 · d 0.50 · cort 0.18 · dopa 0.44 · oxy 0.52 · sero 0.61

derived from conversation + webcam

search — the space lights

shortlist — scored, with and without the lenses

memorysemalexicontaffestrutext onlyfused (all inputs)rank
[m-041] Week-one churn analysis0.920.710.680.340.770.0660.071→
[m-117] Interview — “I didn't know what to do first”0.840.520.630.470.660.0580.064→
[m-006] Onboarding checklist draft0.730.660.490.310.580.0540.061→
[m-238] Fishing with Hank — Stardew's empty first day0.310.080.740.880.710.0120.052↑5
[m-312] Retention benchmark0.690.440.120.050.210.0410.038↓3
[m-301] 1:1 with Keith — compensation note0.220.140.580.400.330.0090.021↑5
[m-092] Landing page A/B notes0.410.330.200.080.180.0230.019↓2
[m-275] Q2 roadmap review notes0.360.220.240.110.190.0190.017↓2
[m-149] Postgres connection pool incident0.290.480.180.090.140.0210.014↓5
[m-261] Hiring loop feedback0.240.170.150.120.110.0140.011↓3
[m-054] Reading note — Norman on affordances0.520.120.090.140.380.0130.010↓3
[m-158] Onboarding email copy v10.330.290.100.060.090.0110.009↓2

Same query. With embodiment, place, time, relationship and state, the shortlist re-ranks — the tangential memory climbs from ninth to fourth.

sieve — task · recency · persona · place · relationship

  • [m-041]Week-one churn analysispassed ✓task fit · recent · product
  • [m-117]Interview — “I didn't know what to do first”passed ✓task fit · recent · product
  • [m-006]Onboarding checklist draftpassed ✓task fit · artifact match
  • [m-238]Fishing with Hank — Stardew's empty first dayretainedtangential — retained (affective 0.88 · structural 0.71)
  • [m-312]Retention benchmarkdroppedlow task fit
  • [m-301]1:1 with Keith — compensation notedroppedsealed · other compartment — scored, never emitted
  • [m-092]Landing page A/B notesdroppedstale
  • [m-275]Q2 roadmap review notesdroppedlow fit
  • [m-149]Postgres connection pool incidentdroppedwrong compartment
  • [m-261]Hiring loop feedbackdroppedlow fit
  • [m-054]Reading note — Norman on affordancesdroppedlow task fit
  • [m-158]Onboarding email copy v1droppedstale

emitted — four memories, in the order they passed

m-0412026-09-11
productopenpinned
weight0.81
recalls34

Week-one churn analysis: 62% of new accounts never return after the first sessio

content
Week-one churn analysis: 62% of new accounts never return after the first session.
context
retention review · shared with Keith · product compartment
reflection
The drop is concentrated in the first ten minutes, not spread across the week.
tangent
Reminded me of a theatre with the house lights up and no one on stage.
cort
dopa
oxy
sero

v -0.32 · a 0.48 · d 0.40

  • m-3120.58
  • m-0060.64
  • m-1170.82
m-1172026-08-27
productopen
weight0.74
recalls21

Interview, week-one churned user: “I didn't know what to do first. So I closed i

content
Interview, week-one churned user: “I didn't know what to do first. So I closed it.”
context
user interview · 2026-08-27 · product compartment
reflection
They did not judge the product and leave; they waited to be told and nothing told them.
tangent
The same sentence appeared in two other interviews that month.
cort
dopa
oxy
sero

v -0.18 · a 0.31 · d 0.46

  • m-2380.55
  • m-0060.61
m-0062026-07-14
productopen
weight0.58
recalls12

Draft of the onboarding checklist — five tasks, no state, no reward for finishin

content
Draft of the onboarding checklist — five tasks, no state, no reward for finishing.
context
design draft · with Mira · product compartment
reflection
A checklist that cannot show progress reads as advertising rather than instruction.
tangent
Progress bars are promises; ours made none.
cort
dopa
oxy
sero

v 0.05 · a 0.27 · d 0.52

  • m-1580.40
m-2382026-06-21
personalopentangentialimage · attached
weight0.44
recalls6
Attached to this memory: two figures fishing from a dock at dusk.meHank

Fishing with Hank at dusk. He talked about Stardew Valley: the first day you do

content
Fishing with Hank at dusk. He talked about Stardew Valley: the first day you do almost nothing, and that's why you come back.
context
evening · the dock · shared with Keith later · personal compartment
reflection
The empty first hour was the retention mechanism, not a failure of content.
tangent
An empty state can invite rather than instruct — signal in the empty space.
cort
dopa
oxy
sero

v 0.71 · a 0.14 · d 0.55

  • m-1170.55
  • m-0540.49

synthesis

Week-one churn isn't a missing-feature problem — it's a silence problem. The dashboard is empty and says nothing m-041, so people do not know what to do first m-117. The checklist helps m-006, but the stronger pattern is the one from an empty evening at a pond m-238: nothing was caught and it still felt good, because the scene never stopped emitting small honest signals. Onboarding should do the same — every empty state should signal progress before the first real result arrives.

creative leap: m-238 → pattern “signal in the empty state” — not derivable from m-041/m-117/m-006

feedback — usefulness written back

memoryused in answerusefulness (agent-scored)Δweight
[m-041]yes0.80+0.10
[m-117]yes0.70+0.09
[m-006]yes0.50+0.07
[m-238]yes0.95+0.12
co-occurrence edgem-238 ↔ m-041 created0.30

The agent scores every emitted memory for how useful it was to this problem and writes the score back. The tangential one that produced the solution scored highest.

the graph changed

The next recall runs on a graph shaped by this one.

A simulated session, eight stages. Every record is an entry in the same store the figures above draw from; hover a memory in either graph to read it whole.