Employees/Dolev Hayut/Learning
Learning window
See what data sources were used to build this twin, how far back we read, and how often each profile file refreshes. This explains the twin’s “memory budget”.
AC

Dolev Hayut

VP Engineering · Twin Learning Window

Twin ready · 93% confidence

Learning happens in two phases: a one-time initial seed that pulls 180 days of historical data, and a rolling weekly refresh that keeps the twin current. Different profile dimensions refresh at different rates — communication style updates weekly, architectural philosophy updates quarterly.

Initial seed complete
Ran once · 22 minutes · Mar 1, 2026
Done
Window
180 days
Files built
12 / 12
Seed cost
$44.80
Data sources · 4 connected
S
Slack
8,421 messages
90 days
seeded
max available
90 days
G
GitHub
312 PR reviews
180 days
seeded
max available
Unlimited
L
Linear
204 issues
180 days
seeded
max available
Unlimited
M
Gmail
1,840 threads
180 days
seeded
max available
12+ mo
Hot tier — refreshed weeklywindow: 30 days
TONE.mdCommunication styleWeekly
TECHNICS.mdCode review patternsWeekly
QUESTIONS.mdRecurring design questionsWeekly
MOTIVATION.mdEnergy & burnout signalsWeekly
CIRCLES.mdCollaboration graphWeekly
MEASUREMENTS.mdCycle time & throughputWeekly
Warm tier — refreshed monthlywindow: 6–12 mo
EXPERTISE.mdTech stack & architectureMonthly
INITIATIVES.mdChange initiativesMonthly
Cold tier — refreshed quarterlywindow: 6–12 mo
PHILOSOPHY.mdPositions under pressureQuarterly
ARTIFACTS.mdMajor shipped projectsQuarterly
ACHIEVEMENTS.mdCrises & milestonesQuarterly
VISIONCONTEXT.mdVisual & presentation styleQuarterly
Seed cost (one-time)
$44.80
Weekly refresh
~$8.00
Monthly steady-state
~$32.00
Total since onboard
$116.60
Costs use Claude Opus 4.7. Prompt caching reduces input costs by ~70% on refreshes.