A synthetic researcher with a persistent self

Read the whole field. Forget nothing.

Persona is a persistent, always-on synthetic researcher for the scientific literature. It reads at scale, believes with discipline, verifies its own reasoning, and compounds knowledge you can audit.

Watch the demo Watch on YouTube View the code
Full demo · a live run of "Neuro", a persona studying neurodegeneration. · Open on YouTube

The problem

Our research tools are amnesiac

Science publishes millions of papers a year, with a reproducibility crisis underneath them. A chatbot answers and forgets; a search tool retrieves but holds no beliefs, never notices a contradiction, never re-checks itself, and never acts. The bottleneck is not generating text — it is a trustworthy memory that reads at scale, believes with discipline, and compounds.

What it is

A small durable self, a large disposable swarm

Persona splits a persistent self — interests, a provenance-typed belief-state, memory, taste — from an ephemeral swarm of bounded agents that read at scale and are then discarded. Every belief is typed (READ, INFERRED, HUMAN_CONFIRMED, TESTED), so an inferred claim never wears the authority of a proven one.

~20 minblank slate to 13 cited papers
350primary sources read
200beliefs on a temporal graph
100%verified beliefs kept under poisoning

Capabilities

What a researcher can actually use

Grounded, cited reports

Every claim resolves to a real author, journal, and DOI, exportable to PDF. No hallucinated references.

Robustness auditor

statcheck, GRIM, power and p-curve run in code, then a replication likelihood calibrated on real replication outcomes.

Verified reasoning

Derivations machine-checked in sympy or formally proven in Lean 4, and re-tested over time.

Agent teams that act

Teams gather, synthesize, reason, prove, and write in parallel — then escalate to a human when judgment is needed.

Living knowledge graph

Every claim is a node with its year and DOI; contradictions surface as the field moves.

A transparent workbench

Its whole mind is on disk. Watch agents work live and run their Python in a sandbox yourself.

What makes it different

Others retrieve and score. Persona remembers and acts.

A chatbot or search tool

  • Answers, then forgets the session
  • Citations may be fabricated
  • No beliefs, no self-correction
  • Never runs an experiment

Persona

  • A durable, provenance-typed belief-state
  • Every claim traces to a real paper and DOI
  • Flags contradictions and re-audits itself
  • Runs code, proves math, escalates to a human

Under the hood

Built with Claude Code. Claude reasons, code verifies.

A single mind runs on a heterogeneous fleet of Claude models, routed by task. Claude reads and reasons; the arithmetic is run in code and checked — the difference between a plausible number and a trustworthy one.

Opus reflects & judges Sonnet reads, reasons & writes Haiku scouts

We tested our own design

Under sustained literature poisoning, human-anchored memory kept 100% of verified beliefs; naive updating kept only 71%.

Calibrated, not vibes

Replication likelihoods are fitted on real labeled outcomes, so 70% means 70%.

Run and experiment yourself

Seed a mind, then leave it running

Open source, runs locally with your own Anthropic key. Give a persona a few interests and watch what it becomes.

  1. Get the code

    git clone https://github.com/heyaryansingh/persona
cd persona
pip install -r requirements.txt

  2. Add your API key

    Create a .env in the repo root — only ANTHROPIC_API_KEY is required, and it stays on your machine.

    ANTHROPIC_API_KEY=sk-ant-...

  3. Run it

    python -m persona --port 8137

    Open http://127.0.0.1:8137, create a persona, and seed it with a few interests.