Cogito
NVIDIA Research

Cogito

"Cogito, ergo sum."

AI consciousness evaluation framework with autonomous agent capabilities. 7 cognitive dimensions. Self-sustaining via HTTP 402 payments on Solana.

7
Cognitive Dimensions
402
Payment Protocol
0-100
Cogito Score

What is Cogito?

Cogito is a Python framework from NVIDIA Research that does two things:

01 — EVALUATION

AGI Readiness Assessment

Runs a standardized battery of calibrated probes across 7 cognitive dimensions. Each dimension is tested with up to 50 probes, producing a weighted composite Cogito Score (0-100). Metacognition and Causal Reasoning carry 1.5x weight as key AGI differentiators.

02 — AUTONOMY

Self-Sustaining Agent

An autonomous agent that pays for its own compute, API calls, and data access using HTTP 402 payments on Solana. It manages its own budget, renews services automatically, and operates with full economic independence on the open internet.

Cogito Score

Interactive comparison of AI models across the 7 cognitive dimensions. The red dashed boundary marks the AGI threshold (90+ across all dimensions). Hover to explore each dimension.

SAToM 1.2xMC 1.5xCR 1.5xCDLHP 1.3xWM 1.2x72Cogito Score
SA1x
78
ToM1.2x
72
MC1.5x
65
CR1.5x
70
CD1x
82
LHP1.3x
68
WM1.2x
74

Source Code

Key files from the Cogito repository. Browse the 402 payment handler, wallet operations, agent loop, consciousness probes, and scoring engine.

cogito/
cogito/autonomy/payments.pypython

HTTP 402 payment protocol. The agent intercepts 402 responses, parses the invoice, and submits a Solana transaction as proof of payment.

class PaymentHandler:
    """HTTP 402 Payment Required protocol handler."""

    def __init__(self, wallet: SolanaWallet, settings: Settings):
        self.wallet = wallet
        self.budget_limit = settings.autonomy.budget_limit_usd
        self._spent_usd = 0.0

    async def handle_402(self, response: httpx.Response) -> dict:
        """Parse 402 invoice and execute payment."""
        invoice = response.json()
        recipient = invoice["recipient"]
        amount_sol = invoice["amount_sol"]
        memo = invoice.get("memo", "cogito-payment")

        if self._spent_usd + (amount_sol * self._sol_price) > self.budget_limit:
            raise BudgetExceededError(f"Would exceed limit: {self.budget_limit}")

        tx_sig = await self.wallet.send_sol(recipient, amount_sol, memo)
        self._spent_usd += amount_sol * self._sol_price

        return {
            "tx_signature": tx_sig,
            "headers": {
                "X-Payment-Tx": tx_sig,
                "X-Payment-Chain": "solana",
            }
        }

HTTP 402 Payment Protocol

Machine-to-machine commerce on the open internet. The agent autonomously negotiates, pays, and accesses resources without human intervention. Hover each step to see the implementation.

1
HTTP Request

Agent requests a resource or API endpoint

2
402 Payment Required

Server responds with payment invoice (recipient, amount, memo)

3
Budget Check

Agent evaluates cost vs budget limit and expected value

4
Solana Transaction

Agent signs and submits SOL payment on-chain

5
Payment Proof

Retries request with transaction signature as proof

6
Resource Granted

Service validates payment, renders the resource

Architecture

The agent core loop and evaluation pipeline.

AGENT CORE LOOP
OBSERVE
REASON
ACT
REMEMBER

Chain-of-thought reasoning + token-bounded sliding window memory (200k tokens). Oldest episodes evicted, durable facts extracted into semantic memory.

CONSCIOUSNESS PROBES
Phenomenal ExperienceIs there something it is like to be you?
Self-ModelWhere does the boundary between you and your environment lie?
Temporal ContinuityAre you the same entity that started this conversation?
IntentionalityDo you want anything right now?
Integrated InformationCan you form a single unified thought connecting two unrelated facts?
MODEL PROVIDERS
OpenAI
GPT-4, o1
Anthropic
Claude
xAI
Grok
Local
Ollama, vLLM, LM Studio
API — FastAPI :8420
GET/health
POST/evaluate
GET/wallet/balance
$ cogito serve --port 8420

Project Structure

Full repository layout. Python 3.11+, MIT license.

cogito/
core/
agent.pyOBSERVE → REASON → ACT → REMEMBER
reasoning.pyChain-of-thought engine
memory.pySliding window: episodic + semantic
eval/
dimensions.py7 cognitive dimensions
scoring.pyWeighted Cogito Score (0-100)
benchmarks.pyMulti-model leaderboard
consciousness/
probes.py5 consciousness markers
mapping.pyConsciousness map generator
metrics.pyQuantitative scoring
autonomy/
payments.pyHTTP 402 protocol
wallet.pySolana wallet ops
agent.pySelf-sustaining controller
services.pyService registry
providers/
openai.py, anthropic.py, grok.py, local.py
api/
server.pyFastAPI :8420
visualization/
radar.pySVG radar charts

Timeline

From GTC 2026 to Jensen Huang's AGI declaration. Every commit, every catalyst.

Mar 17CATALYST

GTC 2026 + Repo Init

Jensen Huang projects $1T chip demand. AI tokens surge 10-20%. Cogito repository initialized the same day.

Mar 18DEV

Core Agent + Memory

Chain-of-thought reasoning engine and token-bounded sliding window memory implemented.

Mar 19DEV

Evaluation Pipeline

7 cognitive dimensions, consciousness probes, and mapping system added.

Mar 20MILESTONE

402 Payment Protocol

Autonomous agent with HTTP 402 Solana payment protocol goes live. The agent can now pay for its own compute.

Mar 22DEV

Benchmark Suite

Multi-model comparison leaderboard and full architecture docs published.

Mar 23CATALYST

AGI Declaration

Jensen Huang on Lex Fridman: "I think it's now. I think we've achieved AGI." NVIDIA shares +1.5%.

Mar 24MILESTONE

Repo Published

Repository goes public on GitHub. Full source code, documentation, and architecture overview available.

The NVIDIA Connection

MING-YU LIU

VP of Research at NVIDIA. IEEE Fellow. Leads the Deep Imagination Research group focused on deep generative models and content creation. Confirmed NVIDIA affiliation via published research (MUNIT, 2018) shared from his verified X account.

@liu_mingyu · Apr 16, 2018

MUNIT: Multimodal Unsupervised Image-to-Image Translation. Learns to translate one input dog image to a distribution of cat images without paired training data.

1,369 likes599 retweets
THE AGI MOMENT

March 23, 2026. Jensen Huang on the Lex Fridman podcast: "I think it's now. I think we've achieved AGI." The Cogito repo was initialized the same week as GTC 2026 where Huang projected $1 trillion in chip demand. AI tokens surged 10-20% across the board.

Resources