how cosmos works

Cosmos is a personal knowledge graph that fills from what you already use and stays at your account.

what cosmos is

Every observation, every connection, every piece of imported context lands in one place that you own. You can read it yourself on your dashboard, talk to it in chat, and let any tool that speaks MCP (Claude Desktop, Cursor, Codex, Zed, the cosmos browser extension, others) read and write through a key you mint and can revoke.

The graph stays portable. Anything you have put into it you can export at any time as a single JSON file. Nothing about Cosmos is meant to lock your data in.

what goes in

Three streams of input, separately tagged so you can always tell them apart.

• Other platforms. When you connect Gmail, Notion, GitHub, your calendar, your browser, Readwise, iMessage, or import a memory dump from Claude or ChatGPT, those services hand Cosmos a steady stream of pages, threads, events, and messages. Cosmos turns each one into nodes in your graph and tags them by where they came from.
• You, directly. Anything you type to Cosmos in chat, or capture by hand, or assert as a self-report. These count as your own voice in the graph and are tagged that way.
• Cosmos's own reflections. Cosmos reads across what you have put in and writes observations back. Patterns it has noticed, connections it has drawn, distillations of longer documents. These are tagged as inferred. They originate from your inputs, but Cosmos shaped the wording.

what cosmos does with it

• Sorts every piece into one of six parts of your life. Work and craft, relationships, health and body, ideas and curiosity, place and logistics, self and inner life. The sort is deterministic for every node, not a fresh guess each day.
• Notices patterns across the graph. Cosmos consolidates recent activity into a small handful of stable observations every twenty hours. It also looks for connections between nodes that share neighbors, and surfaces the parts of your graph that are thin and would benefit from more grounding.
• Answers questions grounded in what you have already put in. When you ask Cosmos something in chat or through a connected tool, the answer is built from your graph. Nothing is asserted without something in the graph to back it.

what you see on /me

Your dashboard surfaces what Cosmos has noticed in language a person can read. Each view is doing one specific thing.

• The graph itself. Every node and every connection Cosmos holds for you, rendered as a 3D map. Filter by life area or by node type. Click any node to see what it is and where it came from.
• Shape. Six axes, one per part of your life. Two shapes overlay each other. The cool solid one is the share of each area that came in through other platforms. The warm dashed one is everything else, what you have said to Cosmos directly plus what Cosmos has reflected back from your own inputs. The two always fill the same circle. Where the cool one is bigger, platforms hold more of you in that area than your own voice does.
• Recent observations. What Cosmos has noticed across your captures and connectors in the last while. Patterns, not facts.
• Where the data came from. A breakdown by source, so the origins are always visible at a glance.

what you can do

• Edit or remove anything. Click any node, change its name, change its content, or archive it. The correction sticks across every later view.
• Connect or disconnect any platform. /connectors shows what is wired and what is not. Disconnect at any time. Revoking access stops future writes but does not undo past ones; deletion is separate.
• Export the whole graph. Through any MCP key, or through /api/polarity/export. JSON, full content, fully portable.
• Take it with you. Anything that speaks MCP can read and write into your Cosmos using a key you control.

how to trust what cosmos shows

Three things back the shapes you see.

• Provenance. Every node carries the source it came from, the date it was first seen, and where applicable the document it was distilled from. The chain is computable from the data itself, not a story Cosmos tells about it.
• Reproducibility. The math behind every dashboard view is deterministic. The shape view is exact weight shares per domain. The categorizer is a single LLM pass per node, persisted, not re-rolled each load. The same input produces the same view.
• Corrections persist. When you edit, archive, or override a node, Cosmos respects that across every later surface. It does not quietly re-add things you took out. Every correction is logged so the system can be retrained against your judgments later.

what cosmos does not do

• Does not tell you what to do. Cosmos surfaces what it has noticed and lets you draw conclusions. No scores, no goal trackers, no nudges toward a better version of yourself.
• Does not rank you. No streaks, no leaderboards, no leveling up.
• Does not share what is yours. Your graph stays at your account. Other platforms only see your data if you authorize a write back. Cosmos does not sell your data, does not advertise against it, and does not feed it into general-purpose training corpora.

For the details of what Cosmos collects, where it lives, and how to delete it, see privacy. For the legal frame, see terms. For what shipped recently and when, see the changelog.