Listens, marks a moment, runs the lookup — while you're still talking.
A free local gate marks the turns worth a second look, a small model on your machine triages them, and your own CLI does the retrieval. Whatever comes back lands on a glanceable panel.
A local rule pass reads the conversation as it arrives and marks the turns worth a second look.
~13% of turns · 65 candidates/hour on 19.4 min of synthetic seed
A small model on your own hardware returns constrained JSON: does this moment deserve enrichment?
llama3.2 3B · no network · no quota · M1 Max / 32 GB
For the moments that pass, KnowMessenger drives the Claude Code or Codex CLI you already have installed. The codex exec median was timed once and that run was not kept; the claude -p one was.
median per engine on 19.4 min of synthetic seed · M1 Max / 32 GB · your account
Context arrives during the conversation — not in the recap.
It listens, marks a moment, retrieves, and surfaces a card. The whole design is about timing: the retrieval runs during the conversation rather than after it.
Listens where you are
Jamie runs on the machine in the room. Capture, transcription, and the whole first half of the decision happen on your device.
Notices, then retrieves
A moment gets marked, a local model decides whether it's worth enriching, and only then does anything heavier run.
Surfaces, doesn't interrupt
The result lands as a small card on a glanceable panel. You look, or you don't. The conversation keeps going either way.
Three stages, and the only paid one is a subscription you already have.
Most of the pipeline never calls a paid model at all. The part that does runs on your account, through a CLI already sitting on your machine.
A free gate, before anything else runs.
A local heuristic reads the conversation as it arrives and decides which turns are even worth a second look. No model call, no network, no cost — it exists so the expensive stages almost never have to run. The rate below was measured on 19.4 minutes of synthetic seed transcript written for the test harness, so read it as the gate's behaviour on that corpus, not as a forecast for your meetings.
~13% of turns · 65 candidates/hour on 19.4 min of synthetic seed transcript · free
A small model, on your own machine.
Candidates go to a local model that returns constrained JSON: does this moment deserve enrichment? On an M1 Max it answers in about 1.9 seconds for $0.00, without touching the network — so there is no quota to burn through and nothing to meter.
llama3.2 3B · constrained JSON · ~1.9s · $0.00 · M1 Max / 32 GB
Your own CLI, on your own subscription.
For the moments that pass, KnowMessenger drives the Claude Code or Codex CLI already installed on your machine, under your existing plan. How long that call takes depends on which one you point it at: 6.6 seconds at the median through codex exec, 18–21 seconds through claude -p — both timed on the same 19.4 minutes of synthetic seed transcript, on the same M1 Max. The claude -p figure comes from a harness run whose output was kept; codex exec was timed once and that run was not, so hold it more loosely than its neighbour.
codex exec 6.6s (timed once, run not kept) · claude -p 18–21s · median on 19.4 min of synthetic seed transcript · M1 Max / 32 GB · your account
No API key. No resale. No metered middleman.
KnowMessenger orchestrates the call — it never holds credentials for a model provider, and it never sits between you and your own plan. That is the whole reason Jamie ships folded into every tier instead of arriving as an add-on with its own line item.
Included in every tier · not an add-on · not an upsell
Announced, not hidden.
A tool that listens in a room with other people in it has to be straight about that. These are defaults, not settings you have to go find.
Announcement-first
Jamie is meant to be announced. The posture is that everyone in the room knows it's on — not that it's quietly clever.
Diarization off by default
Jamie doesn't try to separate and label who said what unless you turn that on. The default is the least identifying one.
Short retention
Transcripts are stored on KnowMessenger infrastructure. The default policy deletes a session — transcript and cards — once it passes seven days from the last capture into it; renaming it does not restart that clock. Deletion runs while you use Jamie: the app sweeps your expired sessions as it serves your requests. The scheduled sweep that would also reach an account that has gone quiet is deployed but currently switched off, so material there can outlive its window. Extended keeps it 90 days; Keep until I delete keeps it until you say so.
Training off by default
KnowMessenger does not train on your sessions, and the default records that you have not consented to any training use. Stage three still hands text to your own CLI, where your own provider's terms apply.
Desktop only — and trigger precision is unmeasured.
A desktop capability, macOS first.
Jamie ships in the KnowMessenger desktop app. A phone can't do this: no mobile OS lets an app reliably restart background microphone capture, so a phone build would be a promise that breaks the first time the screen locks. We would rather say that plainly than run a waitlist for something the platform doesn't allow.
Desktop app · macOS first · no mobile build claimed
Jamie's trigger precision is unmeasured. The harness run that produced the numbers on this page nominated twenty-one stage-one candidates across 19.4 minutes of synthetic seed transcript; stage two then fired on eight of them with the default Claude model, two with Haiku, and nine with the local 3B. Not one of those firings has been rated, so nobody has scored whether the right moments were caught or the wrong ones skipped, and this page makes no claim about it. What is measured is the pipeline itself: the gate's candidate rate on that corpus, the local model's latency and cost on an M1 Max, the median wall time of the final call per engine on that same corpus and machine, and the fact that audio never leaves the device. Even there the two stage-three numbers are not equal: the claude -p median comes from a harness run whose output was kept, the codex exec one from a single pass that was not, and it ships marked as such. When there is a real number for trigger quality, it will appear here with the method next to it.
The questions worth asking.
Including the one we can't answer yet.
Does it record me?
Jamie listens on the machine in the room. Capture and transcription run on your device, and audio never leaves it — the heuristic gate and the local model both read the transcript in place.
The transcript itself is stored, on KnowMessenger infrastructure, and retention is short by default: the default policy deletes a session's transcript and cards once they pass seven days from the last capture into that session. Renaming a session does not restart that clock. Extended keeps them 90 days, and Keep until I delete keeps them until you remove the session yourself.
Deletion is a sweep, not a per-session timer, and it runs while you use Jamie: the app works through your expired sessions as it serves your requests, removing the transcript, then the cards, then the session record. A scheduled sweep that would also reach accounts which have stopped being used is built and deployed on a six-hour schedule — and it currently ships switched off. Until it is armed, material belonging to an account that goes quiet can outlive its retention window, and we would rather say so than let you assume otherwise. Individual cards carry a database expiry as a backstop; session records do not.
Who pays for the AI?
You already do. Stage one is a rule pass with no model call. Stage two runs a small model on your own hardware — about 1.9 seconds on an M1 Max, $0.00, no network. Stage three drives the Claude Code or Codex CLI already installed on your machine, on the subscription you already pay for.
KnowMessenger never holds an API key for a model provider and never resells inference. There is no metered middleman to mark anything up, which is why Jamie ships folded into every tier instead of arriving as an add-on.
Does my audio go to the cloud?
The audio does not. Capture and transcription run on your device and the recording is never uploaded — that is a property of how the pipeline is built, not a setting.
The text does. Be clear-eyed about this: your transcript is stored on KnowMessenger infrastructure. Each session's transcript is written to a KnowMessenger S3 bucket under your own account's prefix, and the session record keeps a short preview of the transcript text so the review list has something to show. That stored copy is what the retention setting governs, and deleting a session deletes it.
Stage three is a separate path. KnowMessenger hands your own CLI a prompt, and that CLI sends it wherever it already sends your prompts, under your own account and your own provider terms — KnowMessenger does not proxy that call and never holds an API key for it.
What about the other person's consent?
Announcement-first is the posture, and the defaults are built to make announcing it the easy path rather than an extra step: diarization — separating and labelling who said what — is off by default, retention is short, and the training default records that you have not consented to any training use. KnowMessenger does not train on your sessions; the announcement itself is still yours to make out loud, since nothing in the product says it for you.
Recording and consent law varies by jurisdiction, and you are the one in the room. Jamie's job is to make the honest version of this easy; the call is still yours to make.
How good is it at knowing when to surface something?
We don't know yet, and we are not going to guess in public. Trigger precision is unmeasured. The harness run nominated twenty-one stage-one candidates across 19.4 minutes of synthetic seed transcript; stage two fired on eight of them with the default Claude model, two with Haiku, and nine with the local 3B — and not one of those firings has been rated, so nobody has scored whether the right moments were caught or the wrong ones skipped.
What is measured is the pipeline: the gate's candidate rate on that corpus, the local model's latency and cost on an M1 Max, and the median wall time of the final call per engine on that same corpus and machine. Even those two engine numbers are not equally evidenced — the claude -p median came out of a harness run whose output was kept, the codex exec one was timed once and that run was not, which is why it ships marked. When there is a real number for trigger quality, it goes on this page with the method next to it.
Does it work on my phone?
No, and it isn't a roadmap item we're going to dangle. No mobile OS lets an app reliably restart background microphone capture, so a phone build would be a promise that breaks the first time the screen locks.
Jamie is a desktop capability, in the KnowMessenger desktop app, macOS first.
Is it an add-on?
No. Jamie is folded into every tier. Because the expensive stage runs on your own subscription and the cheap stages run on your own machine, there is nothing left to meter separately.
Your machine. Your subscription. Your conversation.
Jamie is folded into every tier — not an add-on, not an upsell.