Understand the live situation
The current message is only one clue. Recent messages, the active task, people, application, tools, and immediate work establish what is happening now.
How cognition works
A whole Mind.
A narrow moment.
A useful mind carries far more than it should say or act on at once. Cognition is the process that understands the live situation, admits the right kinds of mental material, ranks that smaller set, and uses only what should govern this moment.
The human inspiration
You do not search your autobiography before answering every question. The conversation already in progress gives you a live situation. That situation cues a small set of beliefs, values, goals, commitments, preferences, procedures, lessons, tasks, and working thoughts. Attention then decides which of those matter most. Only when you need a date, quote, or missing detail do you deliberately search your memory for evidence.
Narrow the full Mind
The situation admits a bounded set from the Mind. This is broader than remembered events and broader than beliefs alone.
beliefsvaluesgoalscommitmentspreferencesprocedureslessonstasksworking thoughts
Rank what was admitted
Within that smaller set, attention weighs fit with the situation, salience, recency, confidence, goals, affect, protection, and what is already active.
Search Memory when needed
Exact history stays separate. Semantic retrieval can find related evidence; lexical search can find an exact phrase, identifier, date, or file. Many turns need neither.
Act, trace, and learn
The model receives the narrow result, not the whole Mind. The system records what shaped the answer, then uses outcomes and corrections to improve later cognition.
The conversation already supplies live context
Recent messages reach the model directly, and inexpensive Mind priming can add a bounded set without an embedding call. Trivial greetings can skip dynamic cognitive retrieval entirely.
Situation-aware admission across the whole Mind
Mind retrieval and episodic Memory search now run as distinct paths. The next runtime phases make candidate admission situation-aware across the whole Mind, prevent the final selector from losing a situation-near item, and add cost budgets and traces.
Watch a Mind change
Experience becomes belief
A customer report is evidence, not truth. The Mind interprets it as a candidate claim, then holds a separate stance: how confident it is, how prominent the claim is, what emotional weight it carries, and whether it is protected.
This chapter (1 of 6)
Notice the differenceThe report and the belief are related, but they are not the same object.
The company's MindEntities are shared across Minds. Beliefs and stances are private to this Mind.
Customers are reporting failed checkouts.Stance 0.60
Salience 0.45
Salience 0.45
The launch should be delayed until the reliability issue is resolved.candidate belief
The launch campaign is already scheduled.Stance 0.55
Salience 0.40
Salience 0.40
Not yet in active attention
Confidence0.00 – 1.00
Salience0.00 – 1.00
Protectionnonenone · user_confirmed · locked
Give this belief your attention.
Salience increases.
Protect this belief.
Much weaker evidence opens conflict.
Introduce contrary evidence.
The threshold decides blend or conflict.
Advance two weeks without attention.
Unprotected salience decays.
Narration
A customer report has become a candidate belief. In this example, the extractor assigned 0.50 confidence. Salience starts at its 0.50 schema default, and the belief is unprotected.
Full simulatorPush several beliefs around at once
The guided walkthrough follows one belief. This lab exposes the whole small Mind: several beliefs, an audit trace, open conflicts, protection, and compressed time.
Try this: let two weeks pass three times and watch what survives. Then recall one belief, protect another, and contradict it.
0days simulated
The trace · what the Mind just did
Everything here maps to a real mechanism: recall reinforces and resets the recency clock, confirmation sets a protection floor, a substantially weaker contradiction against a protected belief opens a conflict instead of overwriting, and decay archives whatever falls below 0.05 salience unprotected. Nearby evidence still blends normally. The emotional belief fades slower than you'd expect. That's the affect dampening.
Go as deep as you want
The walkthrough is the path. These are the mechanics underneath.
Each section below keeps the interactive detail: identity boundaries, attention scoring, affect, decay, write paths, lifecycle, retrieval tools, and the parts that are not finished yet.
The two-layer model
A shared world, a private mind
In plain terms: many Minds can refer to the same launch and the same reliability issue. What each Mind believes about them is its own business, and the walls between Minds are structural, not polite.
Entities, the nouns of the world, are shared. But claims and stances are private. Each proposition belongs to exactly one Mind, enforced by the database itself, not a read-time filter. Two Minds can consider the same decision differently, and neither can read the other's copy.
The company's Mind
Launch --should_be_delayed_until--> reliability issue resolvedproposition · owner: company's Mind · its own embedding
confidence 0.92
salience 0.85
source declared
protection user_confirmed
Shared world layer
event · launch
issue · reliability
issue · reliability
Entities are the common ground. Nothing else is shared.
External advisor's Mind
Launch --should_be_delayed_until--> reliability issue resolvedits own proposition · owner: advisor's Mind
confidence 0.55
salience 0.40
source extracted
protection none
Both Minds hold their own record of the same claim, at their own confidence, anchored to the same shared entities.
Prompt assembly is responder-scoped: when an identity answers, the context assembler loads exactly one Mind, its own. And below the runtime rule, a composite foreign key means an evaluation structurally cannot reference another Mind's proposition. There is no cross-Mind path to leak through.
After candidate admission
Prioritization is a score you can hold in your hand
This calculator starts after the live situation has already narrowed the whole Mind to a bounded consideration set. It demonstrates how one admitted item is prioritized. It does not simulate candidate admission, Memory search, or cognition as a whole.
In plain terms: once an item makes the shortlist, it rises by fitting the live situation, being trusted, prominent, recently used, aligned with a goal, emotionally important, or explicitly protected. Move the sliders and watch one candidate compete for the prompt.
The admitted candidate's state
Composite attention score
0.000
theoretical ceiling 1.05 with the theme bonus
relevance ×0.35
confidence ×0.20
effective salience ×0.20
goal ×0.10
recency ×0.10
confirmed +0.05
theme +0.05
show the working
Recall changes the Mind. Retrieval is not a read-only lookup. The winners get reinforced: their reinforcement count increments, their recency clock resets, and the recall itself is logged so "what did this Mind attend to, and when" stays auditable. Remembering something makes it easier to remember next time. That's the point. One tool, introspect, is deliberately exempt and fully read-only.
Notice what the affect sliders do: it's the absolute value of valence that boosts salience. Strongly positive and strongly negative beliefs both rise. Intensity drives attention, not polarity.
How the Mind forgets
Everything fades, unless it's protected
Left alone, beliefs fade. Emotional weight slows the fade, mirroring the attention boost. Protection levels put a floor under what you've confirmed, and a hard stop under what you've locked.
In plain terms: what you never bring up drifts away, what moved you lingers, what you confirmed can fade but never below "still quite sure," and what you locked doesn't move at all. Pick a protection level and watch 180 days pass.
Protection level
Emotional weight
Curve shows salience from a starting value of 0.9, at the schema's default decay rate, dampened by affect: rate × (1 - |valence|·0.5) × (1 - arousal·0.5). A maximally affect-laden belief decays at a quarter of the neutral rate. Timescale is illustrative; the floors, threshold, and dampening formula are the real ones.
Protection also gates writes. If new information is substantially weaker than a protected belief—an existing-confidence advantage strictly greater than 0.15—the system refuses to blend it in and raises a conflict for review. Nearby evidence still uses the normal weighted blend. So a belief you explicitly confirmed can't drift below "quite confident, quite prominent," can't be archived, and can't be silently overwritten by a much weaker contradicting extraction.
How the Mind spends its night
It sleeps, too
A Mind needs protected time to reorganize experience, not merely time when nobody is talking to it. The target architecture uses five functional states inspired by human sleep: Settling, Light sleep, Deep sleep, REM, and Waking.
In plain terms: sleep begins only when unfinished cognitive work meets a safe quiet opportunity. The Mind closes a stable boundary, selects what deserves replay, stabilizes and clears, searches for provisional connections, then validates its work before waking. The report shows only states and outcomes the system actually measured.
| Term | Mechanism |
|---|---|
| Sleep need | Unfinished cognitive work that the sleep mechanisms can actually reduce. Inactivity alone is not need. |
| Sleep opportunity | A safe quiet period with no active foreground turn, a stable episode boundary, and enough time and resources to do useful work. |
| Settling → Light → Deep → REM | Close a stable boundary, select experience for replay, stabilize and clear, then explore governed connections. Light, Deep, and REM may repeat when useful work remains. |
| Waking | Validate candidate insights, quarantine unsupported changes, publish safe work, refresh retrieval, and carry unresolved need forward. |
| Restoration | A measured improvement in later recall, reasoning, retrieval noise, or efficiency. Completing maintenance jobs is not enough. |
How beliefs enter and evolve
Four doors in, one way to generalize
chat extraction
The main path
Conversations move through an episode state machine, get windowed, and each window is extracted independently and idempotently. Re-running a window cannot double-write. In group chats, only your own lines are attributed to you.
believe()
Direct and synchronous
States a belief outright. The claim is resolved or created, the stance is written at the given confidence, and it's marked confirmed and protected immediately. No extraction sweep needed.
observe()
Asynchronous
Persists free text as a durable, rebuild-safe record and queues it for the normal extraction worker. Eventual, not immediate.
constitutional
Seeded identity
An agent's configuration seeds locked, protected beliefs: its authoritative self-model, which ordinary chat statements can't silently overwrite.
A worked example
One belief, two months
Follow a single belief from a throwaway sentence to protected knowledge. Choose any step; the card on the right shows how its state changes. The mechanisms and defaults are code-backed; the elapsed timeline is illustrative.
Day 0 · stated
"Customers are reporting failed checkouts"
The chat's episode moves through the extraction state machine. The extractor emits a proposition with a summary and an embedding, plus an evaluation in the responding agent's Mind. It assigns confidence from the evidence; this example uses 0.50. Salience starts at its 0.50 schema default, with neutral affect.
Week 1 · reinforced
The leadership team asks whether the launch is still safe
The recall scores this belief well: high relevance, decent recency. Recall is not read-only, so the belief is reinforced, its recency clock resets, and the recall is logged.
Usage calibration now works in its favor: a belief that keeps getting recalled scores higher next time.
Week 3 · confirmed
"Should we delay until the issue is resolved?" Yes.
The evaluation becomes confirmed at protection level user_confirmed. The click itself does not manufacture certainty; it sets 0.7 floors that the next decay pass enforces. The belief cannot be archived, earns a +0.05 retrieval bonus, and opens a conflict when substantially weaker evidence collides with it.
Meanwhile · a neighbor fades
The scheduled-campaign assumption quietly archives
A throwaway belief from the same chat, never confirmed and never recalled, decays freely. Once its salience drops below 0.05, it's archived. Nothing protected ever reaches that threshold.
The belief, live
Customers --are_reporting--> failed checkouts
floor 0.70 · user_confirmed
floor 0.70 · user_confirmed
The launch campaign will proceed on its original date
The exposed surface
Tiered recall, deliberate writes
Retrieval comes in tiers, cheapest to deepest, and most of them change the Mind by design: recalling reinforces. The reads that don't are marked. Writes are explicit.
| Tool | Effect | What it does |
|---|---|---|
| prime_mind | read-only | Cheapest tier. A bounded, salience-ranked set for session startup and ambient priming. No embedding call, no commit. |
| recall_mind | reinforces | The default query-focused semantic tier. Returns a formatted context block; the winners get reinforced. |
| search_mind | reinforces | The widest Mind-only tier. It selects distilled beliefs, preferences, values, goals, constraints, patterns, and decisions without searching conversation history or computing counts. |
| conversation_search | read-only | Searches dated conversation evidence when a turn needs an exact recollection or past-session detail. |
| conversation_aggregate | read-only | Counts and groups captured conversation history, including time buckets. |
| introspect | read-only | A query-less snapshot of top beliefs, holdings, and the working set. The one retrieval tool that never reinforces. |
| believe | writes | Asserts one confirmed, protected belief, effective immediately. |
| observe | writes | Persists free text durably; extraction happens later, on the worker's schedule. |
| resolve | writes | Confirms an evaluation, or locks it so decay never touches it. |
| consult_mind | judgment | Not memory CRUD: a decision gate for non-trivial actions. Returns proceed, proceed-with-caution, ask-user, defer, or block. |
| export_mind | portable | The whole Mind leaves as a signed archive or open linked data: beliefs, goals, and memory, never secrets. |
Deterministic list tools (propositions, holdings, entities) and maintenance operations (rebuild, import) exist too; the table shows the surface most people touch.
Not a replacement, a layer
Everything you need, in one place
A Mind is not something you run instead of Memory. They are distinct, complementary layers. Memory preserves your full history, documents, evidence, and exact records. The Mind forms beliefs and judgment from those sources. thinqOS makes both work together as cognition.
Memory keeps the full history
Every conversation and source can remain durable and searchable, with a coverage receipt, so the system only says "you never told me" when it can back that up. The Mind is distilled from this evidence; it does not replace the record.
The Mind learns from evidence
Documents stay retrievable in Memory while supported claims can become beliefs in the Mind. You get both the source and the stance formed from it, linked but never confused as the same object.
Together, they create cognition
Exact records, searchable sources, beliefs, confidence, relationships, boundaries, and attention work as one system without collapsing storage and judgment into one layer.
Memory + Mind = cognition. Memory preserves what can be known. The Mind holds a point of view about it and decides what matters now.
Honest limitations
Real, deployed, and not finished
Because the Mind is a projection over sources preserved in Memory, it can be rebuilt: replayed cheaply from the event log, or re-extracted expensively from the sources. That property is also why the gaps below are fixable rather than fatal. Stated plainly:
Candidate admission is not yet the complete target.
Recent messages already establish live context, and Mind retrieval no longer automatically searches episodic Memory or runs conversation aggregates. The final selector can still drop a situation-near Mind item. The target is one cheap, situation-aware admission pass across the whole Mind before prioritization, with Memory searched only when evidence is needed.
Privacy scoping is transitional.
Beliefs carry an audience list, but an empty list is treated as universal in some code paths, and holdings aren't audience-filtered yet.
Provenance is not fully end-to-end.
The system can usually show why a belief exists, but not always the complete chain from source observation through extraction, prompt, response, and correction.
Rebuilds are heavy.
Reinterpreting a Mind re-runs extraction over its history: real cost, real time, provider rate limits. A production operation, not a refresh button.