Hey Product Hunt, I've spent 20 years in security for global banking, and the newest failure surface I see is memory: agents deployed on systems that hold no verified state. No record of what is currently true, who committed to what, or what superseded what. Meanwhile my own AI sessions had the same disease. A brilliant planning session on Monday was gone by Wednesday. Every morning started with the same ritual: re-explain the project, paste the context doc, watch half the window disappear before the real conversation even began. MrAgentˣ is the fix I wanted. Say "save" at the end of any AI session, forward email to your workspace, drop in your documents: everything compiles into a context map of facts with dates and sources. When something changes, it looks like this: budget: $40k, since Tuesday's email. Supersedes $32k from March 3. Both on file. The old fact isn't deleted. It's closed, with the date as the receipt. And the next session you open (Claude, ChatGPT, anything MCP) starts already knowing where you left off. Not a transcript. Not a summary. State. One MCP connection, every surface. Wherever your Claude already is (the browser side panel, Slack, desktop, mobile), your context map is already there. It's built from the ground up, deliberately not on RAG. Retrieval finds similar text; it was never designed to hold state, and no amount of tuning makes it know which version of a decision is standing. So the intelligence lives at write time, and there are no flat files anywhere: no MEMORY.md, no append-only logs, no prose summaries feeding on their own prior summaries. Everything is atoms: typed, dated, individually addressable facts that supersede in place, with conflicts flagged instead of merged and two stamped tracks so what your emails prove never blurs with what an AI concluded. We tested it in public against a leading AI memory product, with documents and an answer key. The surprise: the memory product was genuinely good at retrieval, and failed exactly where work actually lives: entity continuity, temporal versioning, commitments and payments, conflict resolution. Those are our main focus. We published everything, including four bugs the test caught in our own pipeline: mragentx.com/blog/the-story-test Three things I'm proud of: - Atoms, not flat files: everything compiles into typed atoms (commitments, decisions, tasks, risks, conflicts, sentiment, and 30+ more types), each dated, sourced, and superseded in place when reality changes. 10 or 100,000 documents and sessions: the cost to load your context stays the same. - Two tracks, never blurred: what your emails prove vs what an AI concluded, stamped at write time. - The real cost in AI is re-explaining yourself, and that's the cost this deletes: sessions start warm from compact atoms instead of re-fed history. Most memory lookups are free calls: deterministic reads, no model in the loop, no per-question meter ever. And the same content never costs you twice: forward an email again, re-upload a document, quote old threads: three layers of dedup catch it before any processing spends anything. It runs on the AI plan you already pay for. We're in private beta and inviting in batches this week. Join the waitlist at mragentx.com, and ask me anything: I'll be here all day.

MrAgentˣ

Save in Claude, resume in ChatGPT. Never re-explain.

Save in Claude, resume in ChatGPT. Never re-explain.