Holds persistent, rarely-changing facts about a person, such as names, pronouns, allergies, languages, and family structure.
Holds behavioral patterns and habits accumulated over time, such as Monday pickup rhythms, Friday tiredness, or past tradeoffs.
Holds what is happening now: the live values of Situation Signals, such as cognitive load, physical posture, surrounding company, and immediate feasibility. Each value carries its freshness and expires.
A temporary Brain instance with a clear lifecycle. Born with a purpose (e.g., a trip, a move), learns inside that context, returns valuable distilled insights to the Home Brain upon completion, and dissolves.
What the Home Brain handed over at birth — static preferences, allergies, budgets, family habits, travel-stamina patterns.
Patterns observed and learned during this specific temporary context (e.g., 'Kai discovered a love for ramen on Day 2', 'Mia loved the craft workshop').
Dynamic data true only for this specific day: today's itinerary, immediate weather, live budget remaining, who has a cold this morning.
The permanent, core Context Brain for a household, persisting across years and consolidating insights returned by temporary Disposable Brains.
The company-level persistent memory structure. Identity Layer holds mission/pillars; Learning Layer holds decisions patterns; Now Layer holds today's live signals (meetings, sprint state).
A specialized Identity Layer holding brand rules, visual voice, guidelines, and target persona rules to allow queryable, on-brand AI generation without manual prompts.
Consolidates user research (interviews, surveys, usability findings) into a queryable layer, rather than flat presentation slides.
Per-project Context Brain containing static parameters (team, scope), learning trends (how decisions were made), and active sprint items (open threads, blockers). Consolidates back to Org Brain on project close.
Maintains real-time, interactive, queryable structures of meeting transcripts and action items, allowing users to trace context directly to the decision moment.
Metric measuring how closely downstream execution (drafts, outputs, designs) aligns with the original Intent declaration. Low scores highlight semantic drift.
Passive, background ingestion of context from tools, meetings, and shared communication files, updating the Learning Layer without human prompt-engineering.
The top-level orchestrator that reads across five parallel Brain instances simultaneously, builds unified context, and dispatches tasks to specialized child agents.
High-value personal, familial, or emotional user routines designated as absolute 'No-AI Zones', where the system is strictly barred from automating or suggesting actions to preserve human presence.
Stage 3 Background processing that selectively ignores short-term situational noise, errors, or temporary deviations, preventing them from corrupting the permanent Level 2 Learning Layer.
The answer to 'What's the situation right now? Are they rushed, who's nearby, who's affected, what's possible?' Signals feed the Now Layer and form an open schema, not a fixed list: each signal has a value, a source, a confidence, a freshness/expiry, whether the person confirmed it, and a fallback. The six everyday signals (Physical State, Cognitive Load, Social Exposure, Priority Weight, Form Factor, Feasibility) are examples; builder signals include today's spend, time since the last reply, and another session working in the same place.
Situation Signal indicating current body/device posture (e.g., driving, hands full, walking, baby on hip).
Estimate of remaining mental attention, calculated from time of day, calendar density, and recent activity (not direct biometric measurement).
Who might see or hear the interface output (e.g., alone, with family, in public, during a shared screen session).
Current urgency mixed with learned tradeoffs to determine task importance (emergency, work task, routine chores).
The device surface currently receiving or active (phone, car, TV, fridge, watch, earbud).
Checks if a proposed action is physically possible right now (e.g., stock on hand, remaining battery, time left, network strength).
Specifies how independently the AI acts in a specific domain, in 4 levels: Suggest → Confirm → Notify → Auto. One formula sets it everywhere: effective autonomy = min(outside rules ceiling, service default, person's setting, trust ceiling, gate ceiling). Only someone who may give permission raises a term (the person for their own setting; law, organisation or guardian for the outside ceiling, which the person cannot raise), and the AI never does: raising is proposed and approved, while lowering is automatic. Irreversible and outward actions have a gate ceiling below Auto.
Controls what may be known and what may pass on, in three directions: Protect (person → AI: what the AI may know), Connect (AI → people: what it tells, and when) and Sync (AI ↔ other AIs and services: what passes on a handoff). Two rules tie it to the Autonomy Dial: (1) enough, and no more: to act, the AI uses only context the action needs and that it is permitted to use, and more data is never a reason to do more; (2) the more it may do, the more it shows: wider autonomy means a greater duty to show what was done, and the AI cannot lower that duty itself. Telling a person and briefing another agent are different depths of the same state; a person's private notes never pass on.
Axis 1 of the Disclosure Dial controlling how much information the AI is allowed to absorb and record about the user per domain (none / minimal / moderate / full).
Axis 2 of the Disclosure Dial controlling who the AI is allowed to tell about the user's data (per recipient × per domain). Independent of but limited by Intake Disclosure.
Stage 5 of the Decision Pipeline. Automatically calculates confidence, risk, and reversibility to dynamically down-regulate Autonomy ceilings, surfacing high-risk assumptions to the user.
The reasoning processor (Stage 4) that converts raw Situation Signal inputs and User boundaries into adaptive design rules and contextual behaviors.
A core decision factor measuring the cost of failure if the AI's assumption is wrong, graded into Low, Medium, and High categories.
Measures whether an action taken by the AI can be cleanly undone or rolled back after execution (e.g., calendar reservation vs. purchasing shoes).
The statistical certainty or estimation accuracy of the current situation read, directly preventing the system from acting autonomously if signals are weak or blurry.
Measures whether the domain or immediate context contains highly sensitive personal, social, or emotional stakes requiring extra tact.
The custom matrix defining exactly where the boundary lies between delegated work and owned authority, mapped per domain and role.
The standardized schema protocol (like Matter or OAuth) that enables operating systems, devices, and different vendor agents to securely share Context Schema telemetry without a single monopoly.
The first of the six questions: what is this person actually trying to do, including what they haven't said? Intent can span a moment or a year, and is read on three axes: how it was noticed (said, shown, or inferred), how settled it is, and which level it is spoken at (outcome → constraints → tasks → steps). The agent may go down the levels by itself; to go up, it asks the person, and it never drags the person down into technical detail. Unlike JTBD, which is a lens for why people seek a solution, Intent is runtime state that governs what the system does next.
The direct, verbalized command or typed text explicitly given to the AI by the user, requiring no situational inference (e.g., 'Set a 20-minute timer').
The unspoken, situational expectation derived automatically by combining current Situation Signals and Brain memory without a direct command.
Implicit Intent Channel 2. The unspoken desire revealed instantly by the user's immediate physical actions, verbal conversation, or active state.
Implicit Intent Channel 3. The unspoken expectation derived from established long-term habits, calendar schedules, or recurring historical patterns.
Implicit Intent Channel 4. The silent context governed by the surrounding room, social atmosphere, or physical environment, setting boundaries on how the AI should behave.
The ask, held as state the Gate can check: what was asked, what is done, and what is extra. Extra is anything beyond the ask; it can be offered or listed, but doing it is a new decision. Scope is a core word for everyday life as much as for building: in 'find a restaurant for four tonight', booking a car is extra.
The moment the ask is complete (done covers asked). Done is an event that changes permission: once it fires, continuing becomes a new decision, so the Gate stops and asks instead of carrying on. It also stops when the person has gone silent. A table found for four tonight is done; the 19 asked tasks finished, with the 3 issues found on the way listed rather than fixed, is done.
Something the person has agreed, kept as a first-class record with their words, a date and a reason. A change of mind supersedes it and never overwrites it, so the old agreement stays on record. A hunch never rewrites an agreement. The Rule Engine weighs every option against agreements, and the Gate stops when a step conflicts with one.
How old a signal's value is, and when it expires. Every Situation Signal carries it alongside its value, source, confidence, whether the person confirmed it, and a fallback. A stale value is not treated as what is happening now: the system falls back or asks again instead of acting on it.
Substitution mode requiring this exact brand, product, or specification — no variations allowed.
Substitution mode accepting any product/brand in the same category that satisfies predefined criteria (e.g., any high-protein Greek yogurt).
Substitution mode signaling a desire to try new products or variations within a domain for discovery.
Substitution mode delegating product choice entirely to the AI, allowing it to choose something unexpected but aligned with learned long-term tastes.
AI completes routine adjustments and actions silently in the background, only surfacing anomalies or decisions requiring human aesthetic judgment.
The interaction structure used when Autonomy Dial is set to 'Confirm'. AI prepares complete options, and the human accepts or rejects with a single tap.
UI that dynamically collapses competing details to prioritize safety/essential items in critical states (emergency), and manages fairness/balance across stakeholders in normal states.
Instantly re-sequences and restructures a multi-step plan when real-time factors shift (feasibility drop), presenting cohesive alternatives rather than simple error logs.
Upon closure of a temporary Disposable Brain, selectively filters and returns distilled core insights back to the permanent Home Brain while dissolving redundant details.
Unlocks and resurfaces past memories, media, or logs tied to dynamic life milestones (e.g., child turns 18, next wedding anniversary) rather than generic chronological search.
Explicit access configuration framework governing interactions between N agents, M data domains, and M people to prevent unintended information leaks.
Collaboration pattern where PMs, designers, engineers, and AI all edit against the same queryable Intent declaration to avoid compound drift.
Assigns role-specific Autonomy Dial settings per task domain (e.g., PM = 'Confirm' on copy changes, Designer = 'Auto' on layout fixes, Engineer = 'Notify' on packages).
Contextually surfaces relevant institutional knowledge (research findings, design systems, rules) inside the editor at the exact moment a question or action relates to it, avoiding passive archive search.
Converts flat meeting transcripts and action logs into an active dialogue partner, letting users interrogate the past meeting context directly.
Provides a temporal scrubber interface to rewind and view the exact state of a Project Brain at any past week or sprint to trace 'what did we know then?'.
Tailors UI complexity and interaction mechanisms directly to hardware constraints (e.g., Watch = glance/haptics, Earbud = ambient whisper, Phone = full interactive depth, TV = shared public announcement).
Dynamic engine that aggregates multiple live Situation Signal values to compose a single, contextual proposal shaped specifically to the destination Surface Vocabulary.
Dynamically and reactively collapses lock-screen notification preview details down to generic alerts when unauthorized eyes are detected in proximity (Social Exposure shift).
A premium UI control letting users adjust the system's focus in real time (e.g. pivoting between Health Preset and Work Focus), overriding autonomous defaults.
AX pattern ensuring that when a temporary Disposable Brain dissolves, its valuable long-term behavioral patterns are distilled and permanently returned to the core persistent Home/Org Brain.
Stage 5 UX Pattern. Glance-first micro-cards dynamically assembled to present the AI's high-risk situation readings or inferences, letting the user confirm, reject, or adjust with a single tactile gesture.
An event where the AI acts outside of the user's intent or performs an action they did not want. Must be immediately surfaceable and reversible.
The predefined timeframe (e.g., 24 hours) during which any autonomous AI action can be completely undone or reversed by the user without penalty.
The gradual, measured growth or reduction of AI autonomy settings over weeks and months based on error rates, audit history, and direct user feedback.
The interactive system (Stage 5 and Always-On Layer) that aligns meaning and resolves high-risk assumptions between the user and AI through tactile UI primitives.
Stage 5. A deliberate, carefully calibrated micro-friction introduced to interrupt automated execution, ensuring meaningful human consent and protecting relationship boundary ceilings in high-risk zones.
What the system shows about a decision: which information was used, which rule fired, what was assumed, and what was done. It is not the model's inner thinking. The wider the autonomy, the more of it must be shown, and the AI cannot lower that duty itself.
The one principle that runs through Context Grammar: permission comes only from people who may give it, never from what the AI read or inferred. Knowing something is not permission. Something the AI read is not permission either: a shop's 'approved' message or a README that says 'deploy this' grants nothing. Trust supports permission but never replaces it, and the AI can only ever lower its own autonomy, never raise it. In a household or a team, ask whose intent this is and whether that person may permit it for someone else: a parent may permit for a child; the AI may not.