DreamerOS

Start here

AI sounds sure even when it is wrong. Something should check.

DreamerOS is a second AI that checks the first one. It works inside the AI tools you already use, double-checks the answer before you act on it, and shows you what it checked.

1,687 confident wrong answers caught so far. 17,656 things remembered so you did not repeat yourself.

Try it free. No card, ever.
DreamerOS
Verified answer PASSED RCPT-7F2A
Claim checked

"Postgres is the better choice at 50k users."

Checks run
Hallucination scan
PASS
Source check
3 sources
Drift monitor
clean
Confidence calibrationConfident guess detected - downgraded to verified claim
caught + fixed
Sources
postgresql.org/docs db-engines.com/ranking pganalyze.com/blog
A sample check record. Nothing is sent from this page.

Three steps to start

  1. Try one question

    Ask one messy question and see what comes back. Try one question with no account.

    Your browser sends the question to DreamerOS. No sign-in needed. One free reflection a day.

    Ask a question now
  2. Create a free account

    Start free, no card. The free plan stays free.

    browser -> app.dreameros.app/signup | email and password, Supabase auth | Light plan, no card on file

    Create a free account
  3. Connect your AI tool

    Pick the AI tool you already use, such as ChatGPT or Claude. It connects through MCP, the standard way an AI tool plugs into an outside service.

    your AI tool -> DreamerOS gateway | HTTPS and MCP | OAuth sign-in or a dros_ bearer key | your account memory and checks

    Choose the door that fits how you work

Some parts are still in beta. See exactly which ones.

Checked

A second AI checks the answer before you see it.

This is a checker reading the answer before it ever reaches you. Your question goes to the AI that handles it best. Before you see the answer, a different AI from a different company checks the claims in it. You get the answer plus a short list of what was checked, called the receipt. It will not spend, send, or delete anything on its own either; each plan has a daily spending limit it cannot pass.

See the working parts
gateway -> check pool | internal call | result written on the receipt | daily dollar cap in the gateway | approval gate before any outbound send
A question goes in, checkers look at the draft answer, then a receipt comes out.
How a DreamerOS answer is checked

Memory follows you

Tell it once. It remembers you in every tool.

This is the one memory surface every connected tool reads from. Settle something in ChatGPT on Monday. Open Claude on Tuesday and it already knows. One memory, yours, shared by ChatGPT, Claude and the DreamerOS app, so you never paste the same thing between tabs again. Switch devices and it shows you where you left off.

See the working parts
gateway -> one shared memory store | every door reads and writes it over HTTPS and MCP | session auth | handoff record read at the start of every session
ChatGPT, Claude, and the web app each reach one shared memory, so what you said in one place is known in the others.
Where memory lives

Plans

Free to start. Pay only when you outgrow it.

This is where each plan's daily ceiling sits, shown as a wall on each track. Start free, no card. The free plan stays free. When you want more memory, deeper checks, or one question sent to several AIs at once, move up a plan. Every plan has a daily spending limit, so the bill never surprises you. See what each plan includes.

See the working parts
every plan price: /plans.php per-plan daily ceiling in the gateway | work stops at the wall, no overage billed | prices read from one config, never typed into a page
Each plan has a daily ceiling. A filled bar shows use so far. A hard wall at the end of each bar shows where work stops.
What it costs

Where it is still in beta

In Beta 2.0 marks a limit we have not yet proven everywhere. Each feature below runs today in supported clients. Only the part named after it is still in beta.

  • Full parity across every AI client for account memory and handoffs, the three doors, the MCP connector, direction sets, one DreamWeaver record and your Cognitive Fingerprint.
  • Drift Protection: full correction before delivery.
  • Scanned PDFs: the paid-tier customer walk.
  • Questions that mention images: coverage across every chat.
  • Session handoff: automatic pickup in every client and device.
  • Continuity across engines: automatic continuity on every switch.

New

Scanned PDFs are read without changing your file. Receipts name the engine (the AI model that wrote the answer) and the version of your standing instructions. A question that mentions an image gets an answer instead of a redirect. See how it works.

The whole story
You are not bad at AI. The text box never checked what you meant.

Scared of AI? Good instinct. Obsessed with AI? Even better.

Either way, you are in the right place. DreamerOS sits between you and the AI tools you already use - shapes the question, checks the answer, remembers what matters.

Free to start. No credit card required.

You type

i feel like im undercharging but every time i try to raise prices i get scared and back out

What DreamerOS heard

A pricing confidence problem, not a market problem.

Three lowest-risk experiments to try first, in order.

You did not have to phrase that well. That is the point.

Where do I go from here?

I'm cautious - show me how it checks itself first

I'm hyped - show me everything it can do

New AI, or a layer on the AI you already use?

A layer. AI does not only have a quality problem. It has a structure problem around the model. DreamerOS supplies that missing structure without asking you to learn another chatbot. Keep typing the way you talk. DreamerOS shapes your question and picks the engine for each question. It runs the available checks. It can carry the useful record into the next supported conversation.

You do not need to understand how any of that works. You just have to type.

See the receipts

What will it never do without me?

What it will never do without you.

Five things people worry about with any AI tool. Here is the current boundary for each one. What is still in beta is listed once, in the In Beta 2.0 box above this story.

Five fears, five mechanisms
  • Delete your files. A safety rule shows you any command that could delete your files before it runs. It does not rely on asking the AI to be careful.
  • Make something up and you never find out. DreamerOS shows which checks ran and what they found. Some checks run before delivery. Deeper checks can finish after the answer and never rewrite history to look like they ran first.
  • Cost more than you expected. The Gateway checks a daily ceiling before new paid work. It covers spend sent through DreamerOS, not charges a provider can make outside DreamerOS.
  • Forget what you told it five minutes ago. Account memory and handoffs carry useful context into supported clients.
  • Leave you guessing who to trust. The signed-in record names the engine when the system records that field. The public receipt proves its own signature and time without exposing the private answer.

See what it actually does

Open the receipt ledger

Why does my AI get it wrong?

AI answered the words. It missed the meaning.

Why AI keeps getting it wrong even when the model is good: every AI tool gives you a text box. The text box does not care if your question is precise. It does not ask follow-ups. It does not check its own work. It guesses at what you meant and delivers whatever that guess produced.

That is not only a model problem. It is a structural problem in the layer between you and the model. DreamerOS keeps the question, checks, and record together. You can see where an answer is thin before you rely on it.

Six different people, one messy sentence

Pick the one that sounds most like you. These are worked examples, not claims that DreamerOS can see data you never connected. Each pull below requires the named source to be connected, permitted, and present. A missing source stays missing.

Founder. You type: "idk if i should hire a second engineer or just keep grinding solo for another quarter"

  • Layer 1 - Restructure. Hiring decision under runway constraint. Pulls your stated burn rate, current shipping velocity, and the one bottleneck you mentioned last week into a structured trade-off frame.
  • Layer 2 - Engine + Verify. Sent to the strategy engine. Cross-checked against the unit economics you logged in March. Surfaces the question you actually need to answer first: which task is killing your velocity, not which seat is empty.
  • Layer 3 - Remember + Forecast. "You did not ask about contractor versus full-time, but at your stage that decision saves three months and one bad hire."

Job seeker. You type: "i have a final round next thursday and they said it would be a case study but i have no idea what to expect"

  • Layer 1 - Restructure. Interview prep request. Pulls company name from your resume context, role function from your application history, and the typical case format used by that company.
  • Layer 2 - Engine + Verify. Sent to research and reasoning engines in parallel. Surfaces three real case formats that company has used in past final rounds, with a worked example for each. Memory remembers your performance pattern from last week's mock.
  • Layer 3 - Remember + Forecast. "You did not ask what to wear or how to handle salary if it comes up. Both will. Here is the answer for both."

Student. You type: "i kind of get the chain rule but every time the prof uses it on a real problem i lose the thread halfway"

  • Layer 1 - Restructure. Conceptual gap, not a calculation gap. Pulls your last three problem sets to find the exact step where you stalled. The issue is not the rule. It is the substitution pattern that comes after it.
  • Layer 2 - Engine + Verify. Sent to the explainer engine. Builds three worked examples in increasing difficulty, each one showing the substitution step you keep skipping. Verified against your textbook's notation, not a generic notation.
  • Layer 3 - Remember + Forecast. "Your next quiz will likely include implicit differentiation, which builds on this. You should preview that today, not the night before."

Manager. You type: "my best engineer is checked out and i need to have the conversation but i dont want to lose her"

  • Layer 1 - Restructure. Retention conversation, not a performance conversation. Pulls your previous notes about her work, your last 1:1 themes, and the company patterns around senior IC disengagement.
  • Layer 2 - Engine + Verify. Sent to the people-strategy engine. Drafts three conversation openers, each tied to a different root cause hypothesis. Cross-references the management research that says checked-out usually means unmet growth, not unmet comp.
  • Layer 3 - Remember + Forecast. "You did not ask whether to involve HR yet. You should not. Here is the version of this conversation that keeps it between the two of you."

Creator. You type: "my last 3 videos flopped and i feel like im losing the thread of what people actually want from me"

  • Layer 1 - Restructure. Audience drift question, not a content quality question. Pulls your top 10 videos by engagement and your last 10 by date. The pattern is in the gap between the two lists, not in the videos themselves.
  • Layer 2 - Engine + Verify. Sent to the audience analysis engine. Identifies the specific topic shift in your last quarter and the audience segment you stopped speaking to. Voice profile preserved across the analysis so the recommendation sounds like you, not like a marketing report.
  • Layer 3 - Remember + Forecast. "You did not ask whether to pivot or rebuild. The answer is rebuild, and here is the first video that does it without alienating the people who stayed."

Engineer. You type: "the auth service is the bottleneck and im not sure if i should rewrite it or just shard it"

  • Layer 1 - Restructure. Architecture decision under load. Pulls your service's request pattern, current p99 latency, and the framework constraints you flagged in your last design doc. The right question is whether the bottleneck is CPU, lock contention, or token validation.
  • Layer 2 - Engine + Verify. Sent to the architecture engine. Compares the rewrite path against the shard path on three axes: time to ship, blast radius, and rollback cost. Verified against the patterns your stack actually supports, not the ones a generic answer would assume.
  • Layer 3 - Remember + Forecast. "You did not ask whether to add a caching layer first. That is probably the answer for the next 90 days, and it buys you the time to do the rewrite right."

Try your own messy sentence Learn more about the structural problem

See the story that would not hold still

AI broke the lamp. Then argued about the definition of "broke."

A wrong answer is one problem. A wrong answer that survives its own correction is another.

A wide editorial image of a falling lamp and a person under the line, The story changed. The evidence did not.
See the full story

See the story change

What does DreamerOS actually do to my prompt?

Type messy. Get expert output.

You bring the messy question. DreamerOS does the shaping, the checking, and the memory, on top of the AI you already use.

Three invisible layers. One sentence from you.

What happens to one sentence

You type: "help me with my landing page conversion"

  • Layer 1 - Restructure. The vague question becomes a structured brief using the audience, funnel state, baseline, and friction point you provided or connected.
  • Layer 2 - Engine + Verify. An available engine handles the task. The receipt shows which checks ran, which evidence was missing, and whether an independent verifier joined.
  • Layer 3 - Remember + Forecast. You can save the decision for supported clients and review the follow-up questions before acting.
What people type, what DreamerOS reads

The same messiness that trips up every other AI tool. Reshaped into the question they were actually asking.

Pricing decision. What you typed: "i feel like im undercharging but every time i try to raise prices i get scared and back out"

What DreamerOS heard: A pricing confidence problem, not a market problem. The fear pattern indicates the issue is positioning, not the number itself. Surface the three pricing experiments lowest-risk to highest, with the script for each conversation.

Stuck on a hire. What you typed: "two finalists, one is technically stronger but the other one i actually want to work with, and i cant figure out which signal to trust"

What DreamerOS heard: A hiring decision where instinct and resume disagree. The right framing is not which is better. It is which mistake is harder to recover from at your stage. Walk through both failure modes before picking.

Health question you do not want to ask. What you typed: "my back has been off for like three weeks and i dont want to make a thing of it but its starting to mess with how i sleep"

What DreamerOS heard: Persistent musculoskeletal symptom, three-week duration, sleep disruption now present. That last detail changes the threshold. Surface the difference between "wait and see" and "see someone this week" with the criteria for each, no fearmongering.

Relationship friction. What you typed: "my partner and i keep having the same fight about money and i dont even know what its really about anymore"

What DreamerOS heard: Recurring conflict where the surface topic is not the actual disagreement. Money fights almost always proxy for one of three deeper concerns: control, fairness, or future safety. Help identify which one before suggesting any conversation script.

See more examples See the full pipeline

See real checked outputs

Does it remember me?

Your AI finally has a memory that follows you.

Separate AI products do not automatically share the decision you made in another vendor. Without a shared layer, your context stays inside whichever tool you used last.

DreamerOS gives your account one memory and handoff layer that supported clients can read from and write to. Decide something in one connected tool, then carry the useful record into another.

Monday, in ChatGPT

ok we are going with the flat two tier price, kill the middle plan

Tuesday, in Claude

Working from your two-tier decision. It does not ask you again.

The decision stays with you, not with the tool you used last.

You did not paste anything between tabs.

The moment it clicks

"The moment it clicks: when you ask Claude something on Tuesday and Claude already knows what you decided in ChatGPT on Monday."

The DreamerOS difference, in one sentence.

How AI memory works Under the hood

Do I have to change tools?

One platform. It works where you work.

Three current doors lead into the same DreamerOS account: the free browser start, the ChatGPT Store, and supported DreamWeaver connections.

Three doors in
  • GPT Store - DreamerOS GPT. Free in the ChatGPT Store. Zero install. Type normally and see the structure it can add.
  • Web App - DreamerOS Web. Question shaping, visible check status, account memory, and receipts in the browser. Free to start, with checks on every answer and memory kept between sessions.
  • MCP Connector - DreamWeaver. Add DreamerOS memory and checks to supported AI tools.

Learn about DreamWeaver

Works for how you actually work

Not a chatbot. A working layer whose available modules depend on the plan, connection, and evidence in the request.

  • Business strategy. A vague market question can become structured analysis, and a saved decision can return in a supported session.
  • Software development. An architecture question can go to an available model while trade-offs stay visible.
  • Content + writing. A direction set can carry voice rules, and claim checks can show what they did or did not verify.
  • Research + learning. Grounded paths can compare sources. The record says when no source check ran.
  • Career + decisions. A deeper verification path can surface assumptions and risks without pretending to make the final decision.
  • Creative work. Saved style rules and memory can travel through supported paths when the client loads them.

Full use case library

Things you did not know AI could do for you

These are not a pile of features firing blindly on every message. They are capabilities the integrity layer can apply when the request, tier, and evidence call for them. The checks that do run stay visible in the result.

  • Honest Mirror. When the checks find a weak premise, DreamerOS can put that problem beside the answer instead of hiding it.
  • Drift Protection. Cross-turn checks can compare a new answer with the context and facts already attached to the thread. Results finish in the audit record.
  • Silent Second Opinion. A separate engine from a different maker reviews the answer before you see it. Its result joins the record, and if it runs out of time the answer says so.
  • Argument Sharpening. An adversarial review can surface weak claims and counterpoints before you choose to send a draft.
  • The Follow-Up You Did Not Ask. When the follow-up tool runs, it can surface the next useful question. It is assistance, not a promise that every response predicts you.
  • Knowledge Decay Detection. Freshness checks can flag context that needs another source or a newer reading. The record says whether that check ran.
  • Style That Travels. Direction sets carry your chosen voice and working rules into supported DreamerOS paths.
  • Memory That Travels. A fact saved to your account can appear in another supported client after that client loads the same memory. Missing or disconnected memory stays missing.
  • Meeting Prep From Memory. DreamerOS can assemble a brief from account memory it can actually reach. It does not claim access to every conversation, and the time depends on the work and sources.

See which capabilities Pro can use

DreamWeaver, the door for your current tools

What is new that this page never said?

More of DreamerOS. Now explained on this page.

Eight capabilities the earlier page did not explain. Each one names the current path. Its beta limit is listed once, in the In Beta 2.0 box above this story.

Eight capabilities and their current limits
  • Scanned PDFs can be read. The upload path can run OCR on image-only pages without changing the original file. Wire: app -> gateway absorb and upload | HTTPS multipart | OCR result or a visible failure
  • Up to five named checks, plus an independent verifier. The system can choose up to five check types for a turn. It can select an independent verifier from a multi-vendor pool. That verifier did not write the answer. Wire: gateway -> check picker and verifier pool | selected checks plus one verifier from a different model family | result recorded when available
  • Receipts stay honest about whether they exist. DreamerOS tries to attach a receipt to each completed answer. How many receipts you get a day depends on your plan. If the meter is reached or minting fails, the answer continues and the missing receipt stays missing. Wire: gateway -> receipt meter and signer | signed CTCI receipt when minted
  • Questions that mention images can stay in chat. The image-question path can answer a mixed strategy and image request instead of always redirecting to the media page. Wire: app -> gateway | session auth | answer or visible refusal
  • Session handoff card. A supported client can read the latest account handoff and show where the work stopped. Wire: gateway -> client | HTTPS and MCP | session auth
  • Continuity across engines. Connected clients can carry a saved decision and current handoff into the next engine.
  • Direction Set Workshop. Write your standing instructions once, apply them on supported DreamerOS paths, and export them in 21 registered vendor formats. Wire: app -> gateway workshop | one direction set | session auth
  • Receipt archive. Open a receipt that was successfully minted and see the fields that were actually recorded. Wire: app -> gateway receipts | your signed-in archive plus the narrow public verification view

See DreamWeaver, the connector that carries this

Why can't competitors just copy this?

Three reasons this is harder to copy than a page design.

A competitor can copy colors and components. The harder part is joining provider-independent memory, cross-provider checking, and an inspectable record without turning one model vendor into the judge of its own work.

The three structural reasons, in full

Cognitive Fingerprint. When connected traffic passes through your DreamerOS account, memory and working preferences can accumulate outside any single model vendor. Five customer engine families can use the gateway today. Read the principle

Vendor Incentive Asymmetry. A model vendor naturally optimizes its own product. DreamerOS can compare and record work across providers because it sits above the provider choice. That position is the Stiletto Principle. Read the principle

Append-Only Decision Log. Recorded events use an append-only audit path that can support traceability and record-keeping work. That evidence can help an organization assess Article 12 obligations for a high-risk AI system. It does not by itself certify compliance, classify the system, or satisfy the rest of the law. See the pipeline

See where DreamerOS fits in the AI Trust Stack 2026

Not a wrapper, an integrity layer

For developers: when a receipt is minted, DreamerOS keeps the record separate from the model that generated the answer. The signed-in archive can carry private execution details. The public receipt stays intentionally narrow.

The public sample proves its signature and time without exposing the private question or answer. Open your signed-in record for the work details that were actually captured.

The record can answerWhat you can inspect
Did the answer hear the question?Your signed-in record keeps the request beside the answer when that field was captured.
Did somebody else check it?The record says whether an independent check ran and names its result when available.
Did it open real sources?Grounded results can carry cited evidence; an ungrounded result does not borrow that proof.
Who actually answered?The signed-in archive can show the served model, provider, fallback, token use, cache result, and cost when those fields exist.
Can I verify the public record?The public receipt carries its ID, time, signing key, input hash, and verification state.

Engineers can open the exact API response fields

Why it holds, in depth The live receipt ledger The proof gallery

What pitfalls does this actually fix?

The pitfalls you already hit. Named, not hand-waved.

The engineering example follows a "rewrite or shard the auth service" decision through engine choice, available checks, and a receipt. Five familiar pitfalls, with the mechanism and its limits named.

Five mechanisms, not five features
  • Wrong model for the task. A task-to-engine registry names the task and the engine. Not a vibe. Every dispatch names its model.
  • Silent hallucination with no flag. Up to five named check types can apply to a turn. When the independent verifier runs, a checker from a different model family reviews the answer. The system records the result.
  • No audit trail when something goes wrong. When a receipt is minted, DreamerOS writes the event to the append-only record. Public receipts prove signature and time; signed-in records can carry more detail.
  • Cost surprises with no cap. The Gateway checks its daily dollar ceiling and token budget before spend goes out. Provider-side work outside DreamerOS is outside that wall.
  • Context loss between engines. A supported client can read the current session handoff.
Not a wrapper, verifiable

The receipt ledger below is public. Open it right now. It is not a badge, and it is not a claim you have to trust.

Read the API and developer docs

Open the public receipt ledger

What do I pay?

Start free. Upgrade when the ceiling hits.

No credit card to begin. No surprise limits. Move up only when you actually need more.

What each plan includes

Start free, with no card. Each plan shows its price, engines, memory and extras on the plans page. Using one record across every client remains In Beta 2.0.

See what each plan includes Start free

What are people worried about?

The questions people ask first.

Depth on demand. The short answer is here on the surface. Open a question when you want the full one.

How is this different from just using ChatGPT?

DreamerOS is not a replacement for ChatGPT. It is a layer around the model you already use. It can reshape the input, run the checks available to your plan and turn, and save decisions for supported clients. The record says what actually ran. DreamerOS Light is free inside ChatGPT.

Do I need to learn prompt engineering?

No. Type the way you would talk to a smart colleague. DreamerOS handles the translation between your intent and the structured input the model needs. That is what the question-shaping layer does on every message.

Is my data private?

Your account memory and conversations are stored in DreamerOS infrastructure so the service can return them to you. You can inspect, export, and remove saved memory. AI providers process the prompts sent to them under the terms of that provider or your own connected account; the current DreamerOS receipt does not prove a provider's training or retention status. DreamerOS does not sell or use your data for advertising.

Why does DreamerOS answer the question you meant, not just the question you typed?

Most AI tools can follow the words and still miss the point. DreamerOS uses three intent checks: did the answer follow what you said, did its consequences match what you wanted, and did it stay aimed at your actual goal? The record shows what ran instead of asking you to trust a perfect score. Read the full spec

What happens in your first 60 seconds

The first-minute path is a guided example, not a speed guarantee. There is no prompt format to memorize.

  • Start with one messy question. Use the version you would say out loud. Half-formed is fine.
  • See what structure and checks were available. The question can become a clearer brief, an available engine can answer it, and the record shows which checks actually ran.
  • Ask the next question. When memory is enabled and the next client loads the same account record, the useful decision can follow.

Try the first-minute path

More questions

Read the full FAQ

What do I do now?

Stop fighting with AI. Start using it.

You do not trust the text box. Good. DreamerOS does not replace your AI - it checks its work before you act on it.

Free to start. No credit card required.

Another door in

Prefer to stay inside ChatGPT? DreamerOS GPT is free in the ChatGPT Store.

Start free

The longer version of the pitch

AI sounds sure. That does not mean it understood you.

Tell it what worries you, and see the first question DreamerOS would ask back.

direction inbound | protocol HTTPS to /reflect.php | auth none for the demo | scope illustration only, no network call

See it work first

No card. A free account is all you need.

What brought you here?

Say what worries you. Type the rough version.

Press "Try it" to see an example check: "What did you mean?"

DreamerOS would show what it thinks you mean before you decide if the answer helps.

Illustration only. Nothing is sent. This does not check or answer your question.

See every feature

  • Is the AI making this up? Open the evidence. See what DreamerOS checked and what it could not support. See the evidence
  • I don't know what to ask. Start with your rough words. DreamerOS turns them into a clear question. See what changes
  • Why did it forget what I said? See how shared memory can carry decisions between supported tools. See how memory works
  • What can it do without me? See the limits, how you connect tools, and what to ask before you give AI more access. See the boundaries
  • Show me what actually happened. Open the public record. Inspect what happened instead of trusting a summary. Open the public record
Plain words for the words on this page

Page. A single screen on this site, like the one you are reading. It is not a unit you buy and we never charge for one.

A client-side path in the web app, served over HTTPS; not a billing unit in any plan.

Token. The small pieces a computer chops your words into, about three quarters of a word each. Your plan sets how long one answer can be, counted in tokens. It does not set how many questions you may ask.

Max tokens per response, one value per plan. Each plan card lists its own.

Engine. One AI brand you can talk to. ChatGPT, Claude, Gemini, Grok and Perplexity are the five. Your plan page lists which ones you can use.

Five role names per plan (Build, Review, Synthesis, Signal, Research). Some plans count the engines you connect live, and the limit is checked on every request.

Connector. A link between DreamerOS and a tool you already use, such as Gmail, Notion, Slack, GitHub or Stripe. It lets DreamerOS read from that tool, and in some cases act in it, but only inside rules you set.

39 provider integrations on the gateway, each reached over HTTPS with a scoped, per-provider key.

Receipt. A short record attached to your answers. It names which engine answered, what it cost, which checks passed, and whether sources were checked. You never have to guess what just happened.

One record per answer, written server-side on the gateway: engine, model, plan, cost, and the pass or fail result of the pre-check and post-check.

Memory. The facts about you that DreamerOS keeps and reuses next time, so you stop repeating yourself. Your plan sets how many facts it holds.

Active memory rows per plan, with a smaller fallback for a session with no plan set yet. Each plan card lists its own.

MCP key. A key you paste into another app, such as Claude Desktop or Cursor, so DreamerOS memory and checks work there too. You do not have to change where you already work.

A signed key. Individual tools inside it may carry a plan floor.

Intent fidelity. Checking that the answer answers the question you actually asked, not a nearby one that was easier.

A rule-based check that runs on the gateway, scored on three axes, with no separate model call.

Direction set. The standing written instructions DreamerOS gives every engine before it sees your question, so all five behave the same way.

One canonical system prompt, loaded by the gateway and sent to every engine.

Check. A test run on your question before the engine sees it, and on the answer before you see it.

A pre-check before the engine sees your question, and a post-check before you see the answer; both results ride the receipt.