|
Until Now.
Never start from zero again.
MrAgentˣ holds what your email said and what you worked out with your AI, knows which is current, and never mixes the two.
The Cost Of Forgetting
The client, the budget, the decision from last Tuesday, pasted in one more time. The bill is not the thinking. The bill is the re-explaining, and it grows with every month of history. Yesterday's session knew all of it. This one starts from zero.
Save in Claude. Resume in ChatGPT. Never re-explain.
Your AI is the engine. MrAgentˣ is the live context map of your work: where everything stands, what changed, what is next.
Join the waitlist →State, Not Prose
Emails, files and chats are prose, and prose has no state. It can’t say what is current, what it replaced, or who said it. MrAgentˣ keeps each fact as state, with its source and date, so your AI answers from what is true now.
How It Works
Three steps, start to finish. A double-tap inside a habit you already have.
Keep The Intelligence
As state, not summaries: what is true now, what it replaced, and the message that proves it. Load it into the next AI and never start from zero.
Every AI you use learns a piece of your work. And keeps it.
Claude knows one part, ChatGPT another, your inbox and your feed the rest. None of them sees the whole.
MrAgentˣ turns it the other way. Everything they learn lands in one context map you hold. Not notes: state. Who is who, what changed and when, what was promised, what was paid.
Asking costs the same however large it grows.
Nothing you taught them leaves with a vendor.
Rent the models. Keep the intelligence.
Double-tap ⌘ and what you are reading is clipped, from any browser or app on Mac. On the phone, save it to Reading List or bookmark it; it is on your context map by the time you are back.
Every fact dated. Every change kept, with what it replaced. Every answer with its source. See the test.
Every browser, every app, every AI. The Live Tracker shows every clip and its progress. Pin one and the author card shows what this author has to do with your work.
Real Test · Same Data, Same Questions
Not a retrieval quiz. The story makes memory do the hard things: track the same people over time (entity continuity), version facts as they change (temporal versioning), close promises when evidence arrives (commitments and payments), and hold a dispute without quietly picking a side (conflict resolution).
None of this is retrieval. Retrieval finds what was said. State knows what is true.
The models were never the problem. Paste the full record into Claude or ChatGPT and both score perfect. The failures live in the memory layer underneath, so that is what we tested.
Run it yourself. Takes 5 minutes, before you find out the hard way your memory layer was wrong.
Under the Hood
Load any of the three tracks as your context, based on what you need.
RE: CHAPS transfer
Funds will be credited within 2 hours, no later than 20 June. Mapped as a [COMMITMENT] (open) and a [RISK] (deadline 20 Jun, open), dated records, not a wall of forwarded text.
Who It Is For
The Context Layer
Your work is the territory: what you received and what you worked out with your AI. Your context map is the live view: where everything stands, what changed, and what is next. Queryable by you or your agent.
No inbox connection. Nothing enters your context map without your action. The trust model is architectural, not a policy.
The emails and documents that arrived sit on one track. The decisions, risks, and next steps from your AI sessions sit on another. What MrAgentˣ inferred across both sits on a third. Each atom is stamped by the Context Engine the moment it is written, and the stamp survives every supersession. A year later you can still tell what the emails prove from what an AI concluded.
Every fact carries its lifespan: when it became true, when it stopped, and when it was last checked. Superseded facts close their window and the newer state takes over, history kept. Stale research gets flagged, not served as current. Your AI never reasons from expired truth.
Where MrAgentˣ Fits
Obsidian / Notion
You write it, you link it, you search it. The system decays the week you get busy.
MRAGENTˣ ADDS
Clips from what already happens, a double-tap, nothing watched. Your AI reads it as state, not pages.
ChatGPT / Claude built-in memory
A best-effort summary inside one vendor's model. It fills up, resets, or gets overwritten.
MRAGENTˣ ADDS
Typed records outside any model. Loads into all of them.
Your email inbox
The record exists but it is buried in threads. Every time you need it, your AI re-reads the prose and you pay the tokens again.
MRAGENTˣ ADDS
Understood once at ingest. After that, commitments, decisions, and risks load into any AI as compact atoms, never a re-read.
Memory APIs
Store chunks for developers to build on.
MRAGENTˣ ADDS
Typed state you can resolve, with a receipt, and a finished product on top: files, topics, history, one load.
Pricing
Metered on storage only. Never on tokens, never on loads, never on AI usage.
FREE
PRO
FAQ
MrAgentˣ is the context layer for humans and their agents, the layer that has been missing across every AI tool you already use.
It follows you. Save in Claude, resume in ChatGPT or your own agent, and the new session picks up mid-thought. Your context map lives outside every model, so switching is one load, not a re-explanation. Connecting a new tool takes one real OAuth login, once, not a manual API key workaround. After that, every switch is just a load, not a new connection.
Bring any of them. MrAgentˣ connects to any client that speaks the MCP standard, with a real one-click OAuth connection, not a workaround. Claude connects on its own dedicated path. Everything else, ChatGPT, Hermes, a self-hosted agent like OpenClaw, or a client that doesn't exist yet, authenticates through CIMD (Client ID Metadata Documents), the current MCP spec's preferred connection method, with RFC 7591 Dynamic Client Registration (DCR) as a fallback for clients that haven't adopted CIMD yet. Both are open standards, not something we built one-off per client. That means a brand-new AI client can connect the day it ships, with zero integration work on our side. Most tools in this space still hardcode a short list of "supported" clients and leave you waiting for the next one to be added.
Yes, and this is where provenance earns its keep. What you ingested (emails, documents) is stamped verified. What an AI contributed in a session is stamped by which AI said it: Claude's take and ChatGPT's take live side by side on the same topic, each labeled. Shop a topic around every AI you use, then isolate what advice came from where and decide with the full picture. Your facts never blur into anyone's opinion, including the AI's.
Yes, and it is kept apart from the email on purpose. Load a topic into Claude or ChatGPT, discuss it, save. The discussion is compiled into the same typed atoms as the email (decisions, risks, assumptions, next steps), stamped as an AI session, and kept separate from what the emails prove. No transcripts, no flat files. Everything lands in the same MrAgentˣ file, a container of atoms, not a document, on the same context map, and you choose what to load: the original emails, what you concluded in the AI, or the latest picture across both. Which model you used is recorded on every atom, so you can filter by it later.
Example: a deal email arrives with a long list of conditions. Most are routine, three need thought. You load the email thread into Claude, read the sender's tone, and work out two approaches. Save. Over the next week more emails come in. Load your last Claude discussion to see which approach you chose and why, then load the latest emails to see what the other side actually did. Plan on one side, reality on the other, and the context map always knows which one is the plan and which one is the fact.
Built-in AI memory is a best-effort summary stored inside the model. It forms without you asking, fills up, resets, or gets silently overwritten, and you cannot see or edit the full set. Your context map is the opposite in every mechanical detail. It is made of the same atoms as the rest of your context map, so changes supersede with dates instead of overwriting. It updates when you say so or at key events, never behind your back. And it has a real UI. Read your entries and delete any source. Chat memory remembers you on its terms. Your memory remembers what you decide, on yours.
Yes. That is the entire point. Every session loads your context map as structured tool calls. Your agent knows what was agreed, what is open, who said what, and what changed. Not because it was told. Because the record exists and it can read it.
Compare the way you do it today: you search Obsidian or your notes, copy prose in, and pay the model to re-read and re-understand it, every single session. With MrAgentˣ the understanding is paid once, when the content arrives. After that, remembering is one click or one search term, returning a compact set of typed atoms scoped to a file or topic, at a fraction of the tokens, every session after.
Memory is not one thing. It is three. Corpus knowledge answers "what does this material say." Document compilers and repo wikis live there. User memory answers "what happened." Preference stores, chat memories, and the agent platforms that carry transcripts across model swaps all live there: portable, but still a pile of what was said. MrAgentˣ is built on the third axis, verified state. It answers "what is currently true." Every fact carries its source and date. Changes supersede with a trail instead of overwriting. Guesses are stamped separately from facts, and forecasts expire instead of aging into truth. Try this test on any memory system. Ask it what is true today, and how it knows. MrAgentˣ answers with the fact, its source, and its date. That is the product.
Quietly outdated facts are the most dangerous thing in any memory system: a confidently recalled old answer is worse than no answer. Every fact on your context map has a validity window and a last-verified date. When something is superseded, its window closes and the newer state takes over, with the old record kept. Facts that have passed their end date, or been replaced, are flagged as stale when loaded, and anything not re-verified in 90 days can be listed for review. You always know not just what the record says, but whether it is still alive.
This is where most memory tools quietly fail. Some pile both versions into prose and let you guess which is current. MrAgentˣ does neither. Every item on your context map is a dated record with a state. When a commitment is fulfilled, the record moves to done and a link connects the resolution back to the original. When two positions genuinely conflict, both stay visible and the conflict itself is flagged, because the disagreement is information. Superseded records are kept, never overwritten. You always see what is currently true, with the full history of how it got there.
Yes, it is a knowledge base. The difference is who it is built for. Confluence and Notion are built for people to read, so nobody does. MrAgentˣ is built for your AI to read. It clips from what already happens, compiles it into dated atoms, and when you open Claude or ChatGPT it loads only the slice the conversation needs. That is how a context window stops being a limit: the relevant part of your whole history arrives in one load, and nothing gets re-explained. You never open it to search. Your AI reads it for you.
Completely. Nothing enters your context map without your action. Delete any atom and the context map updates instantly. Add as much as you want without watching the size: a bigger memory does not make a session heavier, because only a bounded, relevant set ever loads. Load selectively for a light session start, then request specific context on demand. You decide what is clipped, what is removed, and what your agent sees. No background sync. No surprises.
Obsidian and Notion store things you manually write. You build the structure, you create the links, you do the search. MrAgentˣ is built for one purpose: giving your AI the right context before every session. It clips from your actual work, the emails and threads you forward and the sessions you save, structures it into your context map, and loads it directly into any AI. No file-opening. No manual linking. Your AI reads the context map. Obsidian cannot do that.
The mechanism is a skill, not a plugin. Say MrAgent plus what you're looking for in Claude or ChatGPT, and the context arrives as atoms. Work. Say MrAgent, update my file, and the session is compiled back onto the same map, stamped as an AI session. The same skill works in every AI client that speaks MCP, so your context map follows you and the AI never needs to be told where you left off.
Yes, and this is the core discipline. MrAgentˣ never guesses: it surfaces what is stored, every atom with its source and date attached. A session starts small: the topic's goal and its most recent atoms, a working spine, not a file dump. Then it loads dynamically: each question pulls in the next group of atoms most relevant to progressing the discussion, and nothing else. Most tools load file after file and make the AI dig. MrAgentˣ loads meaning at atom granularity, which is why it is both more accurate and cheaper. The record stays honest underneath: emails are never edited, and documents update by re-ingesting. Your context window carries only what the conversation earned.
No model, no middleman. Most memory tools route every request through their own servers and an extra AI call before you get an answer. MrAgentˣ retrieval runs on a dedicated context engine: it compares your question with your stored context by meaning and by keyword, with sources and dates attached to every result. Content stays encrypted at rest and is decrypted on our server only when it is returned to you, nothing is re-reasoned, and nothing leaves except the answer. The one optional exception: with your own Claude key connected in Auto mode, a very short query can be expanded before it runs. Intelligence enters only when you want reasoning, and that is your own Claude or ChatGPT, not a proxy. Fast by architecture, not by benchmark.
No. MrAgentˣ never meters or resells AI usage. The reasoning runs on intelligence you already control: your Claude or ChatGPT plan, or your own API key. MrAgentˣ itself is a monthly plan, not a per-token charge. And it lowers your AI bill rather than adding to it. Starting from zero is what makes AI expensive: every session re-reads and re-reasons over the same history. MrAgentˣ answers from the context layer first, so most requests become marginal retrieval instead of full reasoning. Every new discussion enriches your context map, and because loading is selective, cost stays flat while answers get sharper. Never starting from zero is a cost model.
Nothing, and that is the point. Deduplication runs in layers. Every document is fingerprinted at arrival, so a re-forwarded email or re-uploaded file is skipped before any processing runs. Every extracted fact is fingerprinted too, so an exact repeat is never written twice. And when new content genuinely changes a fact the map already knows, the older version is closed and kept, matched on its exact key, and a near-duplicate is flagged for review instead of being silently copied. Your context map stays clean at every layer, and you are never charged for knowledge you already own. The same content never costs twice.
This is the hidden cost of tools like Obsidian long term. At 10,000 to 30,000 documents, every AI search means loading more context, and your token bill grows with your library. MrAgentˣ inverts that. Search returns a bounded set of results per query, so by the time your agent gets an answer it has loaded a fraction of your library and none of the noise, and what is sent to your AI does not grow with your library.
Most agent memory setups are a file. Frameworks like OpenClaw or Hermes read and write memory as a document your agent loads whole. A file has a size, and size has a cost. Load all of it and you spend tokens re-reading old memory to find one fact. Trim it yourself and the old problem returns: manual maintenance, deciding what to keep and what to lose. Your context map has never been a file. It is atoms on your context map, retrieved by search. Save one thing or ten thousand things. A load or search returns only a bounded, relevant set, however much you have saved. Nothing to trim, no size tax for using it. Save everything. Search finds what matters. And memory is not a bolt-on here. It is part of the same context map, built on the same principle, growing from the same doors: AI sessions, emails, documents. One context map, one set of rules, whatever the source.
No. It reads one thing, when you double-tap ⌘, through the accessibility layer. No script runs in the page and nothing is sent to the platform. The phone path fetches public post pages the way a browser would, nothing else.
No. It drafts a reply using your profile, your voice and style, and your context map, so it knows what you have already said to this person. It stays factual: no guesses, nothing added that is not in the post or your context map. You read it and post it. Change a line and it learns your style from the edit, so the next draft is closer. Nothing goes out by itself.
If their handle is in an email signature you already have, it links them. Where the match is only likely, it asks you first.
Context Windows Close. AI Forgets Everything. Your Work Should Never Start From Zero.