Local-first models
Transcription, embedding and reasoning run on your hardware by default. No account is required for memory to work.
Speech, screen, people, commitments — searchable, with the clip or file each fact came from. Playable evidence, not a black-box summary.
Pricing moves to usage-based in Q3.
Standup · 00:12:04 — play clip
You owe Marcus the revised SOC 2 scope.
Screen · contract-v4.pdf, p.3
Beta invites gated until the audit log ships.
Design review · 00:31:57 — play clip
The problem
The decision happened in a call. The context is in a thread. The commitment is in someone's head. A week later nobody can say who agreed to what, or why.
Notes apps store text without provenance. Chatbots answer confidently from nothing you can check. Meeting recorders keep a transcript on someone else's servers and forget you the moment the call ends.
The missing piece isn't summarisation. It's a memory you own, where every fact can be traced back to the moment it was said.
The loop
Speech, screen, files and calendar — recorded locally, segmented, and timestamped as you work. Nothing is uploaded to start.
Signal becomes durable records: a claim, a decision, a commitment — each one linked back to the moment it came from.
Who said it, who owns it, who's waiting. Your working graph assembles itself from what already happened.
Your phone or a colleague's assistant can ask a question. Your policy answers it. Raw memory stays put.
Drafts arrive grounded in memory. Send, buy, mutate — none of it happens until you press approve.
Trust is the product
Transcription, embedding and reasoning run on your hardware by default. No account is required for memory to work.
Every fact keeps a pointer to its source — the audio segment, the screen frame, the file and page. Click it and hear it.
Recall can shape a draft. It can never be the reason something was sent. Those are separate systems on purpose.
Send, buy, schedule, mutate — each one stops at Approve / Edit / Cancel, with the memory it relied on shown alongside.
People & work
Mnemos keeps track of who you actually work with, what's open between you, and which commitments are drifting. It's assembled from evidence you can inspect — never guessed, never inferred from a contact list.
SOC 2 scope revision
Marcus — Legal
Usage pricing model
Dana — Finance
Audit log surface
Priya — Design
Connection without exfiltration
Your phone, or a colleague's assistant, sends a question. Your machine decides — under a policy you wrote — whether to answer, and returns an answer, not the record behind it. Raw memory never leaves the device it was captured on.
Peer asks
"Did we agree the audit log ships before beta?"
Your machine answers
"Yes — decided 14:26, design review."
clip withheld · scope: colleagues · expires 30d
Action, gated
Drafts in the browser, on the desktop, from your phone — all grounded in memory, all stopped at the same gate, with the evidence they leaned on shown next to them.
"Revised SOC 2 scope attached — matches what we agreed on the 14:26 review, minus the vendor annex."
grounded in 2 memories · play both clips
Who it's for
Twelve conversations a day, every one of them load-bearing.
Decisions scattered across calls, docs and three tools.
Long threads of reasoning that have to stay attributable.
People who refuse cloud-memory-by-default, for good reasons.
Honest status
Mnemos is early. It runs locally on Windows first. We'd rather you knew exactly what exists today than find out after installing.
Run it locally
One email when the local build is ready to install. No newsletter, no drip sequence, no third-party analytics on this page — the form posts to our own edge and stores nothing but your address.
Prefer the reasoning first? Read how trust and locality are designed.
We don't track you on our own website.