Algorithm

AI research by Polarity Lab

Pressure tests

Fair opponents only. We compare against thread plus prompt, vendor memory, capture tools, and notes apps.

vs. Thread + prompt

ScenarioThemCosmosVerdict
One hard question, no historyStrong with a good promptLittle added valueParity
Sixth session on the same goalContext rot, re-explainingPath, capture, recall persistWin
What am I forgetting?Cannot knowRecall due from graphWin
Mail + calendar + chat togetherManual paste each timeConnectors, one graphWin

Skeptic: I will write a great system prompt.

Prompts do not survive app switches, do not schedule your forget curve, and do not show drift with receipts weeks later.

vs. ChatGPT / Claude memory

ScenarioThemCosmosVerdict
Preferences in one chat productGoodGood via MCP in same productParity
Memory across Cursor, Claude, and webSiloed per vendorOne account graph, any MCP clientWin
Inspect and correct a memoryLimited visibilityProvenance on nodes, dismiss or editWin
Export and leave vendorWeak or locked.polarity exportWin
Drift readoutNot a product surfaceCore /me loopWin

Skeptic: ChatGPT already remembers me.

Inside one assistant. Cosmos is the map every agent reads, and you own it.

vs. Rewind / Limitless

ScenarioThemCosmosVerdict
What did I do Tuesday?Strong searchConnectors + queryParity
Am I living the life I meant?Not the jobDrift + pathWin
Agent ground in CursorNot primaryMCP nativeWin

Skeptic: Rewind already records my life.

Recording is not authorship. Cosmos proposes moments; you decide what counts, then measures alignment to path.

vs. Notion / Obsidian

ScenarioThemCosmosVerdict
Store what you wroteExcellentCapture + chronicleParity
Read what you did not writeManualConnectors with consentWin
Automatic drift detectionNoneHeading vs. connectorsWin

Skeptic: I already have a second brain.

Second brains store what you bothered to write. Cosmos reads across life surfaces and shows when behavior diverges from heading.