Omnesis Brain

Architecture

From context substrate
to cognitive substrate

The context substrate is the ground truth. It is factual and it does not interpret. But raw retrieval is not understanding. The Omnesis Brain aims to derive the deeper meaning of the things in your life, and to enable intelligence that is proactive rather than reactive.

Collector

Supports 40+ sources

Context substrate

Omnesis materializes your digital life into a searchable people and document graph

Omnesis Brain

Background agent

Reacts to new, edited and deleted data and keeps the derived substrate up to date.

Sweeps

Recipes for pattern finding and periodic reviews on their own cadence. Sweeps can be contributed by the community.

Decision gates

A small, cheap model that decides what is worth a run, how urgent it is, and what is worth keeping.

Derived substrate

Wiki pages

Pages about your projects and topics, with a short root page on top. Every sentence cites its evidence.

Time index

Extract semantic times and ranges buried in your context substrate into a queryable index.

Annotations

Facts about your documents and the people in your life, each backed by a quote.

Loops & tasks

The unresolved things in your life (actions, decisions, commitments), suggested to you as tasks.

Observations

Things the Brain noticed that you probably want to know. They feed your morning brief.

Wikis

The Brain maintains
a wiki about your life

The Brain writes and maintains wiki pages about what is going on in your life: projects, places, people, arrangements. A short root page gives the overview and links to more detailed pages, which can themselves point to even more detailed ones. Underneath, pages refer to the loops the Brain is tracking, and eventually to the raw documents in your sources: emails, chats, files.

A layered graph. The root wiki cites the Home, Finances, Work and Health pages; those cite Kitchen renovation, Q4 launch and Marathon training; Kitchen renovation cites the Studio Northstar page and the Pay deposit loop, which Finances cites too. At the bottom, the largest layer: about twenty overlapping source tiles (WhatsApp chats, Gmail threads, Google Drive files, a calendar event, Notion pages, Strava runs) that the pages and loops cite.

Claims

Every sentence in a
wiki page is a claim

Wiki pages are Markdown, but every bit of text has to sit inside a claim tag. A claim has an ID and references to what supports it: a passage in a source document, a field of a loop, or a claim on another wiki page. A separate verifier model then checks that the references actually entail the claim, taking the page title and section into account, before the claim is marked as verified.

The Kitchen renovation page in raw Markdown. Three claim tags, budget, deposit and start, each wrap a heading or sentence and cite one reference: a passage of the Revised quote email, the state field of the Pay deposit loop, and the crew claim on the Studio Northstar page.

A passage of a document

source:doc#evidence:passage The exact passage in an email, chat, file, etc.

A field of a loop

loop:id#field:state What a loop currently records, e.g. whether it’s done.

A claim on another page

wiki:page#claim:id This is how pages build on each other.

Staying up to date

When a document changes, the
Brain knows what to re-check

Because claims can point at other claims, everything the Brain derives forms a dependency graph that goes all the way down to your source documents. When a document is edited or deleted, the claims that cite it are marked as stale right away (no model call needed), then the claims that cite those claims, and so on up the graph. Links that are only there for navigation don’t propagate anything. The Brain then only re-checks what was marked stale.

The Revised quote email is edited. The Pay deposit loop that cites it, the Kitchen renovation page, the Home page and the root wiki all turn orange and read needs re-check. Studio Northstar, the Health branch and their sources stay green, and a dashed navigation link from Kitchen renovation to Marathon training carries nothing.

Answering

How the agent
uses the wiki

The root wiki is injected in the Omnesis agent’s system prompt, so the agent knows the broad strokes of what is going on in your life from the start of every conversation. From there it can open the pages it needs. If a claim it reads is verified and none of its evidence has changed, the agent can use it as is. If the claim is stale, it goes back and reads the sources.

In practice this means fewer tool calls and fewer tokens, so answers come back faster. This is still very experimental.Experimental

Asked when the kitchen fitting starts, the agent goes from the root wiki in its system prompt to the Kitchen renovation and Studio Northstar pages. A verified, current claim answers directly: week of 16 November. A stale claim sends it down into a mountain of source documents, where three arrows pick out the few it re-reads.

Decision gates

Using a small model to
decide what is worth doing

The Brain is the only part of Omnesis that spends tokens without anyone asking a question, so it has to be careful about what it spends them on. Before the expensive steps, it asks Jev, a small decision model from TypeSafe, for a score. Jev is fast and cheap, every verdict is recorded so you can inspect it, and if Jev is unavailable the Brain falls back to the safe option.

New documentarrives or changes
WorthIs it worth reading at all?
UrgencyNow, within the hour, or later?
Brain agentreads, links, writes
Record checkDoes it add real value?
SalienceDoes it earn your morning?
Morning briefone message a day

Jev also decides whether a change can affect a claim that depends on it, when a page is due for a review, whether a renamed task is still the same task, and whether a suggested task is worth showing you. When I replayed ten weeks of history, the worth gate alone skipped about 58% of email runs, which were mostly noise.

Sweeps

Sweeps: recipes for
recurrent reflection

Some things about your life are only visible from a distance — changes in health, finance or relationship patterns. Sweeps let a background agent step back and look for those patterns, at a cadence you define, and note what it finds for your morning brief.

1 --- 2 name: Day ahead 3 cadence: 1d 4 at: "05:00" 5 --- 6 7 Prepare the user for the day that is starting. Look at what is happening today — meetings, appointments, scheduled events — and for the ones that merit it note an informative observation for the morning brief: what it is, when, with whom, and the context worth having to hand, drawn from the surrounding correspondence. Include the time of the event it concerns and cite the documents you drew it from. Check the open loops falling due today or very shortly and note a reminder observation where it would help. Then look a week out — across both source-owned schedules and your own recorded interpretations of upcoming time — and for anything genuinely approaching that needs preparing, note an observation and open a loop if it is a real obligation nobody is tracking yet. If along the way you find a tracked obligation has already been fulfilled — a receipt, a confirmation, a reply that settles it — close its loop as done with a note and withdraw the observations it made moot; never delete a loop that was actually completed, because the record of what happened is worth keeping. Check existing observations before writing anything, so an earlier pass's observation is updated rather than duplicated. Where a future event would benefit from a fresher picture, schedule yourself a follow-up to recheck it and refresh its observation only if it will still be useful at the next morning brief. Most days only a handful of things merit an observation, and an empty pass is a fine outcome.

sweeps/may-day.md

Morning brief

One morning brief
a day

The Brain doesn’t send you notifications as it goes. When it notices something you probably don’t know yet and would want to act on, it records an observation, with quotes from its sources and a date after which it stops mattering. Every morning, one run goes through the open observations and writes a single message in your Morning brief conversation. If there is nothing worth saying, there is no message.

Tuesday

Good morning. Three things today:

The revised quote is $1,100 higher. Studio Northstar now asks $18,400, and the deposit is due Friday.
Site visit at 10:00 with Maya Reeves. She asked for the final appliance list.
Your passport expires in May; the March trip needs six months of validity.

Based on 3 things

Revised quote · Studio Northstar

Passport renewal reminder

Site visit · 10:00

Add a task to pay the deposit tomorrow

Added: Pay Studio Northstar deposit, planned for tomorrow.

Tasks

Tasks work without the Brain,
and better with it

Tasks are a simple bullet journal that you and the agent can manage. When the Brain is on, the loops it tracks show up as suggested tasks, and the tasks you accept are kept up to date for you.

Without the Brain

Available to everyone

A bullet journal for today, a specific day, a month or someday, on the portal, iOS, Android and widgets.
Ask the agent in any conversation to add, complete, move or strike a task, and undo it.
Tell Omnesis “I sent the invoice” and it gets ticked off, with an undo.

With the Brain

Experimental

Suggested

Register for the Riverside half marathon

AcceptAccept for…Ignore

Send the signed plan to Maya Marked done by Omnesis · Undo

Loops show up as suggestions you can accept, schedule or ignore.
Accepted tasks get marked done when a source says so and the verifier agrees. You can always undo.
If a change touches what you wrote yourself, it shows up as a proposal you confirm or dismiss.
It learns from what you accept and ignore.

Tell Omnesis

Tell the brain,
while on the go

Speak it or type it, in the app, on your watch or with Siri. Each note is read by an agent that acts on it: it can add, complete or move tasks and remember facts about you, and it leaves a receipt you can undo. If it needs to ask you something, it opens a conversation and notifies you. Your note also records when you said it, where you were and what you were doing.

An Apple Watch listening: “Hey Siri, Omnesis Capture — what should I remember?”, with the spoken line “Jameel just said he’s send me the revised quote by Friday”

Watches

Watch Engine: detect conditions
expressed in natural language

Ask Omnesis to be notified when a certain condition is met. Omnesis compiles your condition expressed in natural language into a DSL evaluated by the Watch engine.

Tell me when I receive an email about a property listing and I have not replied in 3 days.

1 { 2 "watch": { 3 "name": "x-listing-unanswered", 4 "nl_query": "Tell me when I receive an email about a property listing and I have not replied in 3 days.", 5 "firing_policy": "stays_active", 6 "ontology_fingerprint": "eef3032ba8488f3e61a908f6b28f5832", 7 "nodes": [ 8 { 9 "id": "listing", 10 "type": "source.document_event", 11 "filter": { 12 "source": "gmail", 13 "event": [ 14 "created" 15 ], 16 "documentType": "email", 17 "people": [ 18 { 19 "role": "sender", 20 "isSelf": false 21 } 22 ] 23 }, 24 "recall": { 25 "semantic": { 26 "query": "property listing, a house or flat for sale or to rent, with an asking price or a viewing", 27 "threshold": 0.35 28 } 29 }, 30 "judge": { 31 "proposition": "This email is about a specific property being offered for sale or to rent — a listing, a viewing, or an agent's details for one — and it is addressed to the recipient rather than a mass mailing", 32 "output_schema": { 33 "why": "string" 34 } 35 }, 36 "output_map": { 37 "doc_id": "$e.docId", 38 "thread_id": "$e.metadata.extra.threadId", 39 "why": "$judge.why" 40 } 41 }, 42 { 43 "id": "my_reply", 44 "type": "source.document_event", 45 "filter": { 46 "source": "gmail", 47 "event": [ 48 "created" 49 ], 50 "documentType": "email", 51 "people": [ 52 { 53 "role": "sender", 54 "isSelf": true 55 } 56 ] 57 }, 58 "output_map": { 59 "thread_id": "$e.metadata.extra.threadId" 60 } 61 }, 62 { 63 "id": "unanswered", 64 "type": "stateful.wait", 65 "inputs": { 66 "listing": { 67 "role": "arm", 68 "key": { 69 "thread_id": ".thread_id" 70 } 71 }, 72 "my_reply": { 73 "role": "cancel", 74 "key": { 75 "thread_id": ".thread_id" 76 } 77 } 78 }, 79 "on_collision": "reset", 80 "duration": "3 days", 81 "output_map": { 82 "doc_id": "$n.listing.doc_id", 83 "why": "$n.listing.why" 84 } 85 } 86 ], 87 "sink": { 88 "input": "unanswered", 89 "output_map": { 90 "evidence": "$n.unanswered.doc_id", 91 "why": "$n.unanswered.why" 92 } 93 }, 94 "delivery": { 95 "kind": "omnesis-notify" 96 } 97 } 98 }

watches/x-listing-unanswered.json

source.document_event listing source=gmail · event=created LLM judgesemantic recall

source.document_event my_reply source=gmail · event=created

stateful.wait unanswered 3 days after arming collision: reset

sink from unanswered Notifies your devices omnesis-notify

(thread_id)

(thread_id)cancels

Tell me when sleep, HRV and heart rate all deviated from my usual on the same day.

1 { 2 "watch": { 3 "name": "x-body-signal-day", 4 "nl_query": "Tell me when sleep, HRV and heart rate all deviated from my usual on the same day.", 5 "firing_policy": "stays_active", 6 "ontology_fingerprint": "eef3032ba8488f3e61a908f6b28f5832", 7 "nodes": [ 8 { 9 "id": "morning_tick", 10 "type": "source.time", 11 "recurring": "0 9 * * *", 12 "output_map": { 13 "due_at": "$e.dueAt" 14 } 15 }, 16 { 17 "id": "sleep_dev", 18 "type": "sql", 19 "inputs": { 20 "morning_tick": { 21 "role": "arm" 22 } 23 }, 24 "query": "WITH daily AS (SELECT cast(start_time AS DATE) AS day, avg(value) AS v FROM health_sleep WHERE metric_slug = 'asleep_minutes' GROUP BY 1), subject AS (SELECT v FROM daily WHERE day = $today - INTERVAL 1 DAY), baseline AS (SELECT avg(v) AS avg_v, count(*) AS n_days FROM daily WHERE day BETWEEN $today - INTERVAL 31 DAY AND $today - INTERVAL 2 DAY) SELECT ($today - INTERVAL 1 DAY) AS day, baseline.n_days >= 14 AND abs(subject.v - baseline.avg_v) >= 0.30 * baseline.avg_v AS fires, round(100.0 * (subject.v - baseline.avg_v) / baseline.avg_v, 1) AS deviation_pct FROM subject, baseline", 25 "output_map": { 26 "day": "day", 27 "deviation_pct": "deviation_pct" 28 } 29 }, 30 { 31 "id": "hrv_dev", 32 "type": "sql", 33 "inputs": { 34 "morning_tick": { 35 "role": "arm" 36 } 37 }, 38 "query": "WITH daily AS (SELECT cast(start_time AS DATE) AS day, avg(value) AS v FROM health_vitals WHERE metric_slug = 'hrv' GROUP BY 1), subject AS (SELECT v FROM daily WHERE day = $today - INTERVAL 1 DAY), baseline AS (SELECT avg(v) AS avg_v, count(*) AS n_days FROM daily WHERE day BETWEEN $today - INTERVAL 31 DAY AND $today - INTERVAL 2 DAY) SELECT ($today - INTERVAL 1 DAY) AS day, baseline.n_days >= 14 AND abs(subject.v - baseline.avg_v) >= 0.45 * baseline.avg_v AS fires, round(100.0 * (subject.v - baseline.avg_v) / baseline.avg_v, 1) AS deviation_pct FROM subject, baseline", 39 "output_map": { 40 "day": "day", 41 "deviation_pct": "deviation_pct" 42 } 43 }, 44 { 45 "id": "hr_dev", 46 "type": "sql", 47 "inputs": { 48 "morning_tick": { 49 "role": "arm" 50 } 51 }, 52 "query": "WITH daily AS (SELECT cast(start_time AS DATE) AS day, avg(value) AS v FROM health_vitals WHERE metric_slug = 'heart_rate' GROUP BY 1), subject AS (SELECT v FROM daily WHERE day = $today - INTERVAL 1 DAY), baseline AS (SELECT avg(v) AS avg_v, count(*) AS n_days FROM daily WHERE day BETWEEN $today - INTERVAL 31 DAY AND $today - INTERVAL 2 DAY) SELECT ($today - INTERVAL 1 DAY) AS day, baseline.n_days >= 14 AND abs(subject.v - baseline.avg_v) >= 0.30 * baseline.avg_v AS fires, round(100.0 * (subject.v - baseline.avg_v) / baseline.avg_v, 1) AS deviation_pct FROM subject, baseline", 53 "output_map": { 54 "day": "day", 55 "deviation_pct": "deviation_pct" 56 } 57 }, 58 { 59 "id": "all_three", 60 "type": "stateful.and", 61 "inputs": { 62 "sleep_dev": { 63 "role": "arm", 64 "key": { 65 "day": ".day" 66 } 67 }, 68 "hrv_dev": { 69 "role": "arm", 70 "key": { 71 "day": ".day" 72 } 73 }, 74 "hr_dev": { 75 "role": "arm", 76 "key": { 77 "day": ".day" 78 } 79 } 80 }, 81 "deadline": "1 day", 82 "on_collision": "ignore", 83 "output_map": { 84 "sleep_pct": "$n.sleep_dev.deviation_pct", 85 "hrv_pct": "$n.hrv_dev.deviation_pct", 86 "hr_pct": "$n.hr_dev.deviation_pct" 87 } 88 } 89 ], 90 "sink": { 91 "input": "all_three", 92 "output_map": { 93 "sleep_pct": "$n.all_three.sleep_pct", 94 "hrv_pct": "$n.all_three.hrv_pct", 95 "hr_pct": "$n.all_three.hr_pct" 96 } 97 }, 98 "delivery": { 99 "kind": "omnesis-notify" 100 } 101 } 102 }

watches/x-body-signal-day.json

source.time morning_tick cron 0 9 * * *

sql sleep_dev DuckDB over your analytics

sql hrv_dev DuckDB over your analytics

sql hr_dev DuckDB over your analytics

stateful.and all_three every arm, within 1 day collision: ignore

sink from all_three Notifies your devices omnesis-notify

(day)

Tell me when an unusually large payment is made and it can't be explained by a bill I received or a subscription.

1 { 2 "watch": { 3 "name": "x-unexplained-large-payment", 4 "nl_query": "Tell me when an unusually large payment is made and it can't be explained by a bill I received or a subscription.", 5 "firing_policy": "stays_active", 6 "ontology_fingerprint": "eef3032ba8488f3e61a908f6b28f5832", 7 "nodes": [ 8 { 9 "id": "spend", 10 "type": "source.analytics_row", 11 "table": "bank_transactions", 12 "op": [ 13 "inserted" 14 ], 15 "predicate": "amount < 0", 16 "output_map": { 17 "txn_key": "$e.row.transaction_key", 18 "amount": "$e.row.amount", 19 "currency": "$e.row.currency", 20 "merchant": "$e.row.merchant_name" 21 } 22 }, 23 { 24 "id": "above_p99", 25 "type": "sql", 26 "inputs": { 27 "spend": { 28 "role": "arm" 29 } 30 }, 31 "query": "WITH baseline AS (SELECT quantile_cont(abs(amount), 0.99) AS p99, count(*) AS n_txns FROM bank_transactions WHERE amount < 0 AND currency = $n.spend.currency AND transaction_key <> $n.spend.txn_key AND transaction_date BETWEEN $today - INTERVAL 30 DAY AND $today) SELECT baseline.n_txns >= 14 AND abs(CAST($n.spend.amount AS DECIMAL(18,4))) > baseline.p99 AS fires, round(baseline.p99, 2) AS p99 FROM baseline", 32 "output_map": { 33 "p99": "p99" 34 } 35 }, 36 { 37 "id": "explained", 38 "type": "llm", 39 "mode": "investigation", 40 "inputs": { 41 "above_p99": { 42 "role": "arm" 43 } 44 }, 45 "proposition": "This charge is explained by something already in the corpus: a bill or invoice received for it, or a recurring subscription with proof of it — a confirmation, a renewal notice, or earlier charges of the same shape from the same merchant.", 46 "output_schema": { 47 "explained": "bool", 48 "proof": "string" 49 }, 50 "fire_when": "NOT explained", 51 "scope": { 52 "horizon": "180 days" 53 }, 54 "budget": { 55 "tool_calls": 12, 56 "tokens": 40000 57 }, 58 "output_map": { 59 "proof": "proof" 60 } 61 } 62 ], 63 "sink": { 64 "input": "explained", 65 "output_map": { 66 "evidence": "$n.spend.txn_key", 67 "merchant": "$n.spend.merchant", 68 "proof": "$n.explained.proof" 69 } 70 }, 71 "delivery": { 72 "kind": "omnesis-notify" 73 } 74 } 75 }

watches/x-unexplained-large-payment.json

source.analytics_row spend table=bank_transactions op=insertedpredicate=amount<0

sql above_p99 DuckDB over your analytics

llm explained deliberates with tools LLM investigationhorizon 180 days

sink from explained Notifies your devices omnesis-notify

Agents

Watches can wake your external agent
and trigger workflows

Your external agent can ask Omnesis to watch for something. When the condition is met Omnesis wakes the agent so it can pull the necessary context (via Omnesis privacy gate) and run its workflow — no polling.

Openclawonline

any time I receive a new property listing by email, crawl it, confirm 2 bedrooms and under 30 min from work, then ping me here09:12

Sure thing, let me ask Omnesis to create a watch!09:12

create watch

Inbox

2-bed flat, Maple Court

Nora T.to me 14:47

hi Maya, sending over one new property that just got into the market…

watch fires

wake agent

read email
crawl listing
check commute
draft message

Openclawonline

Sonia sent you a property listing with 2 beds that’s a 15 min walk from work!14:49

From context substrate to cognitive substrate

Background agent

Sweeps

Decision gates

Wiki pages

Time index

Annotations

Loops & tasks

Observations

The Brain maintains a wiki about your life

Every sentence in a wiki page is a claim

A passage of a document

A field of a loop

A claim on another page

When a document changes, the Brain knows what to re-check

How the agent uses the wiki

Using a small model to decide what is worth doing

Sweeps: recipes for recurrent reflection

One morning brief a day

Tasks work without the Brain, and better with it

Without the Brain

With the Brain

Tell the brain, while on the go

Watch Engine: detect conditions expressed in natural language

Watches can wake your external agent and trigger workflows

From context substrate
to cognitive substrate

The Brain maintains
a wiki about your life

Every sentence in a
wiki page is a claim

When a document changes, the
Brain knows what to re-check

How the agent
uses the wiki

Using a small model to
decide what is worth doing

Sweeps: recipes for
recurrent reflection

One morning brief
a day

Tasks work without the Brain,
and better with it

Tell the brain,
while on the go

Watch Engine: detect conditions
expressed in natural language

Watches can wake your external agent
and trigger workflows