← /daily-startup-idea
June 25, 2026

Bloom

“A Bloom filter for ‘have I told you this story?’ Probabilistic, compact, runs in the inhale before you launch into it. Sometimes silences you about anecdotes you never actually told — a feature, not a bug.”

The Problem

You are 34. You have, generously, eleven stories. You tell the one about Cher at the airport once a quarter to the same six people. They have heard it. They are too polite to say so. Your partner now mouths the punchline into her water glass at a delay of exactly one beat.

The Solution

A small daemon between your brain and your mouth. Every anecdote you tell is hashed three ways; the resulting bits get flipped in a per-listener bit array. Before any story leaves your face, we check the array. All bits set → almost certainly already told.Any bit unset → definitely fresh, proceed.

BLOOM · per-listener bit array · k=3
listener: wife
candidate:“so the time I almost met Cher at LaGuardia —”
h₁(s)
→ bit 0x07
h₂(s)
→ bit 0x12
h₃(s)
→ bit 0x1A
0
0
0
1
0
0
0
1
0
1
0
0
0
0
1
0
0
0
1
0
0
1
0
0
0
0
1
0
0
1
0
1
verdict: ALREADY TOLD (p > 0.98)action: shut up · sip water

Latency

Sub-millisecond at the conversational threshold. The classifier triggers on the leading phrase — “so this is funny —”, “okay wait,”, “did I ever tell you about” — and completes lookup during the inhale. You never know it ran. You just suddenly do not have anything to say.

False-Positive Tuning

Sometimes Bloom says toldwhen you haven’t. You will, occasionally, swallow a perfectly good story about a raccoon. We are at peace with this.

WHISPER
FPR 1% · 8 KB / listenerFor the verbose; ships with apology budget.
STANDARD
FPR 5% · 2 KB / listenerDefault. You lose ~1 raccoon story per month.
AGGRESSIVE
FPR 25% · 256 B / listenerRecommended for relatives. You will be quieter than usual; no one will ask why.
MONK
FPR 75% · 32 B / listenerFilter assumes you’ve told everything. You haven’t. It’s fine.

Repeat Rate by Listener

Anonymized telemetry from beta cohort, normalized over a 90-day window. The cache is, predictably, worst for the people who love you most.

spouse / partner
96%
mom (on speakerphone)
88%
best friend since 6th grade
81%
the group chat
64%
coworker getting coffee
22%
person seated next to you on plane
3%

Pricing

Solo
personal
$8 / mo
One bit array per listener, up to 50 listeners. Stories TTL at 6 months — rotate the rotation.
Couples
most popular
$19 / mo
Shared bit arrays so they don’t hear it from both of you at the dinner party. Includes one weekly “tell her anyway” override.
Family Plan
enterprise
$59 / mo
Up to 6 mouths. Holiday Mode: pre-warmed with every story your dad has ever told. Filter says told. Filter is correct.

TAM

Anecdotes retold globally per year (est.)
4.6B adults × 11 stories × 14 listeners × 9 retells
= 6.4 trillion repeats
Of which, 80% are the Cher one.

The Ask

$3.5M seed. Builds the on-device hashing model, the inhale-trigger classifier, and a graceful-fail mode where the daemon murmurs “maybe” in your earpiece and you, the human, get to decide. Two engineers, one linguist, and a comedian on retainer to label training data.

Series A pitch:“We’re Datadog for the things you’ve already said.”