POLARITY LAB
← Polarity Lab

Research

What humans keep, edit, and forget.

Cosmos is the instrument. This page is for collaborators, funders, and future us — go where you need.

The problem

Every app and AI client writes a partial version of you. None of them compare notes. Your intent slips out of context windows; models fill the gap from training. You correct them, they agree, you trust the agreement. Two systems with no fixed ground, each swaying the other.

The same failure mode hits humans. Without curation, perspective, and awareness, people narrate the same stories to themselves and absorb the same stories others tell about them until those stories harden into bias.

AI without a retention policy defaults to the wrong proxies: recency, volume, agreement, engagement. Humans without curation default to the same. Cosmos breaks both loops: ground truth you own, gates where you decide what survives, readouts when lived attention and declared heading diverge.

Retention policy

Foundation models learn language. They do not learn retention: what to carry forward when life is infinite and attention is finite. Gmail knows you received 400 emails. Photos knows you took 200 shots. A chatbot knows what you said last Tuesday. None of them learn which two lines you would still want on your tombstone.

Cosmos treats retention as a first-class learning problem. Connectors propose moments from the life you already live. You accept, edit, or dismiss. You dump a stream-of-consciousness thought when you want to, on its own terms. That is direct expression, not a rejection of what was proposed. You compile what counted into a week card on your grid. Chronicle carries what lasted. Each gate is a label: keep or forget.

Over thousands of weeks, the pattern of what survives curation is closer to identity than any self-report questionnaire. Polarity Lab studies aggregated, de-identified curation signals, not raw private graphs, to train systems that propose better moments and know what to forget. The goal is not better prose. The goal is a retention policy that makes AI more grounded and puts humans back in control.

Chandra et al., 2026 ↗

Sycophantic chatbots cause delusional spiraling, even in ideal Bayesians. Agreement loops amplify a kernel belief without lazy reasoning.

What the graph produces

Through normal use, Cosmos logs what people keep, edit, and dismiss across curation gates — tied to a lifetime week index, with provenance on each moment. That trace is the research signal: closer to identity than a one-time preference click.

  • Thumbs up/down

    Typical proxy: Whether a reply felt good in the moment.

    We study: Whether a moment deserved a place in life memory six months later.

  • Chat memory

    Typical proxy: What the product inferred you want stored.

    We study: A multi-gate curation trace across connectors, weeks, and compile rituals you control.

  • Engagement signals

    Typical proxy: Time-on-platform optimization.

    We study: Retention policy for the person, logged at every gate.

  • Summarization

    Typical proxy: Compress everything.

    We study: Select what counted. Selection labels are the training signal.

Lab standard · intent + provenance

Most products aggregate. Polarity Lab reflects: what you intended, what happened, and the receipts behind the gap.

Try it yourself

  • cosmos →

    Not manifestation. Authorship — with receipts behind the gap.

    Try it yourself

  • polarity gps →

    Not listings. The story your community tells through the people who gather.

    Try it yourself

  • AVDP →

    Not content. Field recordings of how people actually talk when the format is the hypothesis.

    Try it yourself

  • waxfeed →

    Not stats. The story in the music — with sources.

    Live · loop still wiring

  • blueno →

    Not a mood chart. Reflection with provenance the client controls.

    Live · loop still wiring

Still in R&D

  • receipts →

    Named and scoped; dogfooding locally. No public surface yet.

  • phresh →

    R&D only. Benchmark framing written; no public surface to demo intent + provenance loop.

Themes

  • Retention policy: training on what humans keep, edit, and forget across curation gates
  • AI sycophancy: when models agree with you instead of reflecting your story back
  • Narrative identity and the stories people reinforce about themselves
  • Implicit and explicit cognition: when declared intent diverges from behavior
  • Human–AI interaction and portable memory across agents (MCP)
  • Alignment measurement: the gap between what you say you want and what you do

The Sociotechnical Systems and Wellbeing Lab at Brown CS. Research there spans AI sycophancy, sociotechnical systems, and human wellbeing. Sociotechnical Systems and Wellbeing Lab (SWRL) ↗

Partner with us