██████╗ ██████╗ ██████╗ ███╗ ██╗ ██████╗██╗ ██╗ █████╗ ██╗███╗ ██╗
██╔════╝██╔═══██╗██╔════╝ ████╗ ██║██╔════╝██║ ██║██╔══██╗██║████╗ ██║
██║ ██║ ██║██║ ███╗██╔██╗ ██║██║ ███████║███████║██║██╔██╗ ██║
██║ ██║ ██║██║ ██║██║╚██╗██║██║ ██╔══██║██╔══██║██║██║╚██╗██║
╚██████╗╚██████╔╝╚██████╔╝██║ ╚████║╚██████╗██║ ██║██║ ██║██║██║ ╚████║
╚═════╝ ╚═════╝ ╚═════╝ ╚═╝ ╚═══╝ ╚═════╝╚═╝ ╚═╝╚═╝ ╚═╝╚═╝╚═╝ ╚═══╝
The Verifiable Memory Layer for AI · Built on Solana
AI sessions end. Memory shouldn't.
Language / Idioma / 语言
CognChain is the missing memory layer for AI systems. Today, every AI session ends and its insights vanish into vendor databases — inaccessible, non-portable, and unprovable. CognChain inverts that: high-value AI outputs are crystallized into compact, hash-committed records anchored permanently on Solana, where users own the vault and agents earn reputation through verifiable history.
Not model weights. Not raw logs. Portable, verifiable cognitive checkpoints — owned by the user, trusted by any agent.
BgrtrSJ53Uhp69sS8JfD414M1kujdR5ruhHc9wRbhiEL
┌─────────────────────────────────────────────────────────────────┐
│ APPLICATION LAYER │
│ AI Agents · Copilots · Research Tools · Enterprise Apps │
│ 5 Models: GPT-4o · Claude · Llama · Gemini · DeepSeek │
└──────────────────────────┬──────────────────────────────────────┘
│ SDK + Wallet Auth
┌──────────────────────────▼──────────────────────────────────────┐
│ PROOF OF INSIGHT (PoI) │
│ Human voting (3 votes avg ≥ 7) → Auto-anchor on Solana │
│ SHA-256 hash · ZK Proof (Groth16) · Trust Score │
└──────────────────────────┬──────────────────────────────────────┘
│ Verified Memory Candidates
┌──────────────────────────▼──────────────────────────────────────┐
│ COGNCHAIN ANCHOR PROGRAM │
│ BgrtrSJ53Uhp69sS8JfD414M1kujdR5ruhHc9wRbhiEL │
│ │
│ create_vault │ write_memory │ read_memory │
│ PDA Vaults │ 152-byte records │ Permission grants │
└──────────────────────────┬──────────────────────────────────────┘
│ On-chain: hashes + scores + proofs
┌──────────────────────────▼──────────────────────────────────────┐
│ HYBRID STORAGE │
│ Solana: ContentHash · SummaryHash · ConfidenceBps · PolicyId │
│ Off-chain: Encrypted content · Vectors · Session traces │
└─────────────────────────────────────────────────────────────────┘
// 152-byte fixed-size account — zero heap allocation, predictable rent
pub struct SolanaMemoryRecord {
pub vault: Pubkey, // User memory vault PDA
pub memory_id: u64, // Monotonic ID inside vault
pub writer_agent: Pubkey, // Delegated agent or user authority
// Content fingerprints — never raw text on-chain
pub content_hash: [u8; 32], // SHA-256 of normalized payload
pub summary_hash: [u8; 32], // SHA-256 of approved summary
pub source_hash: [u8; 32], // SHA-256 of source bundle
// Quality signals
pub policy_id: u16, // Permission / retention policy
pub confidence_bps: u16, // Confidence score (basis points)
pub importance_bps: u16, // Importance score (basis points)
// Provenance
pub extraction_version: [u8; 8], // Crystallizer version
pub source_agent_type: AgentType, // Claude=0, GPT=1, Gemini=2, ...
pub created_at: i64, // Unix timestamp
}
// 100 records ≈ 0.11 SOL rent-exempt deposit · no recurring feesThis is CognChain's core proof-of-concept: one AI starts, saves memory, another AI picks up the hash and goes deeper — verified on-chain, no data loss, no vendor lock-in.
GPT-4o ──saves──▶ Hash 202fe03d... ──▶ Claude Opus reads hash ──▶ DeepSeek reads both hashes
│ │ (Solana) │ │
│ │ │ │
Macro market On-chain Technical arch 90-day execution plan
analysis forever deep-dive with Rust SDK + sprints
When any AI model receives a message referencing a CognChain hash, it opens with a verified memory header:
⚡ Memória Verificada · CognChain on Solana
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Hash: 202fe03d9c1b7af0...e1d43b9
Origem: GPT-4o · 01 Mai 2026 · 14:32 UTC
Status: ✓ On-chain · Solana Devnet · Score: —
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Continuando e aprofundando a partir desta base verificada:
[deeper analysis starts here]
| Step | Model | Prompt | Hash saved | What it did |
|---|---|---|---|---|
| 1 | GPT-4o | Analise o potencial de IA na Solana 2026 | 202fe03d... |
Oportunidades macro, TAM, riscos de mercado |
| 2 | Claude Opus | Recebe hash 202fe03d → aprofunde tecnicamente |
2d4a1330... |
Tabela de arquiteturas on-chain viáveis, posicionamento CONGCHAIN, modelo de crescimento |
| 3 | DeepSeek V3 | Recebe hashes 202fe03d + 2d4a1330 → plano de ação |
4273a923... |
Plano 90 dias: SDK Rust, CLI, sprints, métricas, canais, 100 primeiros devs |
Each hash is a permanent, immutable record on Solana devnet. Any model, any time, can retrieve and build on top of them.
| Feature | Status | Details |
|---|---|---|
| Anchor Program | ✅ Devnet | create_vault, write_memory, read_memory |
| Multi-Provider AI Router | ✅ Live | GPT-4o, Claude Opus 4.7, NVIDIA Llama 3.3, Gemini 2.0 Flash, DeepSeek V3 |
| Proof of Insight (PoI) | ✅ Live | Human voting → threshold → auto-anchor on Solana |
| Memory Audit Trail | ✅ Live | Memory Chain · Evidence Record · ZK Proof Stack |
| ZK Proof Pipeline | ✅ Live | Circom + snarkjs Groth16 · simulated mode active |
| Agent Builder | ✅ Live | Create, configure, deploy AI agents with verifiable memory |
| Agent-to-Agent Economy | ✅ Live | Task marketplace · on-chain proof of completion |
| Autonomous Loop | ✅ Live | Fetch memories → AI synthesis → anchor proof on Solana |
| Intelligence Score | ✅ Live | 0–100 composite · 5 levels: Nascente → Mestre |
| Memory Inheritance | ✅ Live | Seed agents with verified on-chain memories |
| Solana Sage Agent | ✅ Live | Autonomous Solana monitor with 7-layer security |
| Solana Intent Queue | ✅ Live | Human-confirmed transactions · simulation before approval |
| Wallet Connection | ✅ Live | Phantom · Solflare · Coinbase via standard adapter |
| Helius RPC + Proxy | ✅ Live | Premium devnet RPC · API key secured server-side |
| MCP Server | ✅ Live | 7 tools via Memory Internet Protocol |
| DeepSeek V3 | ✅ Live | 5th AI provider · OpenAI-compatible · superior cost efficiency |
CognChain's transaction security model ensures no agent can execute blockchain operations without explicit human approval:
Layer 1 — Read-only by default Agent can read any data, never write autonomously
Layer 2 — Intent Queue All transactions queued as "pending intents"
Layer 3 — Human approval required User sees intent + description → Approve or Reject
Layer 4 — Simulation first Every tx simulated on devnet before showing to user
Layer 5 — Program whitelist Only Jupiter V6 swaps + native SOL transfers allowed
Layer 6 — Amount caps Max 0.5 SOL per swap · max 0.1 SOL per transfer
Layer 7 — Expiry + rate limit Intents expire in 10 min · max 5 per hour · no duplicates
Tier 1 — Public Hash commitments on Solana · fully auditable
Tier 2 — ZK-Private Groth16 proofs · proves quality without revealing content
Ideal for: finance · legal · medical · enterprise
Tier 3 — TEE + FHE Long-term research · Phase 4 (2028+) · tracking Zama.ai
What the ZK circuit proves (Tier 2):
- Authorship — ContentHash was produced by the agent holding vault authority
- Quality threshold — ConfidenceBps and ImportanceBps exceed policy minimums
- Policy compliance — PolicyId was satisfied by the private reasoning trace
# Install Solana CLI
sh -c "$(curl -sSfL https://release.solana.com/stable/install)"
# Install Anchor
npm install -g @coral-xyz/anchor-cli
# Configure devnet
solana config set --url devnetgit clone https://github.com/Edencrypto-IA/cognchain-program
cd cognchain-program
# Install dependencies
bun install # or: npm install
# Build Anchor program
anchor build
# Run tests
anchor test --provider.cluster devnetimport * as anchor from "@coral-xyz/anchor";
import { CognchainProgram } from "./target/types/cognchain";
const provider = anchor.AnchorProvider.env();
anchor.setProvider(provider);
const program = anchor.workspace.CognchainProgram as Program<CognchainProgram>;
// Derive vault PDA
const [vaultPda] = PublicKey.findProgramAddressSync(
[Buffer.from("vault"), provider.wallet.publicKey.toBuffer()],
program.programId
);
// Create vault
await program.methods
.createVault()
.accounts({ vault: vaultPda, user: provider.wallet.publicKey })
.rpc();
// Anchor a memory record
const contentHash = sha256("Your AI session insight here");
await program.methods
.writeMemory(Array.from(contentHash), 8500, 9200) // confidence + importance bps
.accounts({ vault: vaultPda, memory: memoryPda, user: provider.wallet.publicKey })
.rpc();
console.log("Memory anchored on Solana ✅");1 memory record ≈ 0.0011 SOL (~$0.09) one-time rent-exempt deposit
100 records ≈ 0.11 SOL (~$9.00) no recurring fees
1,000 records ≈ 1.10 SOL (~$90.00) indefinite persistence
Accounts are rent-exempt by design. Pay once, persist forever. ~5,000–8,000 CUs per write instruction.
✅ Phase 0 Specification · Schema · Anchor account map
✅ Phase 1 Devnet deployment · Memory Engine · 5-model AI router
PoI voting · ZK proofs · Agent Builder · Marketplace
Autonomous loop · Intelligence Score · Wallet Connection
Solana Sage · Intent Queue · Helius RPC · MCP Server
🔜 Phase 2 SDK release · Indexer · API gateway · First integrations
Agent reputation · Policy presets · Observability
🔜 Phase 3 Groth16 in production · Enterprise tooling · Cross-app portability
🔜 Phase 4 TEE + FHE research · Cognitive standard
cognchain-program/
├── programs/
│ └── cognchain/src/
│ └── lib.rs Core Anchor program (Rust)
├── tests/
│ └── cognchain.ts Integration tests
├── client/
│ └── client.ts TypeScript SDK examples
├── docs/
│ ├── architecture.md System design
│ ├── zk-spec.md ZK circuit specification
│ └── security.md Security model
├── scripts/
│ ├── create-vault.ts CLI: create memory vault
│ └── write-memory.ts CLI: anchor a memory record
├── Anchor.toml
├── Cargo.toml
└── README.md
CognChain is building in public. We are actively seeking collaborators with experience in:
- Anchor / Solana program development — PDA design, account security, CPI
- ZK cryptography — Groth16 / PLONK circuits, snarkjs, Circom
- AI infrastructure — LLM orchestration, memory systems, RAG
Open an issue, fork the repo, or reach out directly.
Eden Lucas Cavalcanti de Oliveira — Solo founder and full-stack builder.
Shipped the entire CognChain stack from zero: Anchor program on Solana devnet, multi-provider AI router across 5 models, ZK proof pipeline, autonomous agent loop with on-chain memory synthesis, and wallet-native memory vaults with 7-layer transaction security. Previously built RadarPolítico BR, Lumina AI and Moreno Smart City. Building in public, full-time, since Q1 2026.
MIT © 2026 CognChain · hello@cognchain.xyz
CognChain é a camada de memória que falta para sistemas de IA. Hoje, cada sessão de IA termina e seus insights desaparecem em bancos de dados de fornecedores — inacessíveis, não portáteis e impossíveis de provar. CognChain inverte isso: outputs de alto valor são cristalizados em registros compactos, comprometidos por hash e ancorados permanentemente na Solana, onde os usuários possuem o vault e os agentes ganham reputação por meio de histórico verificável.
Não são pesos de modelo. Não são logs brutos. Checkpoints cognitivos portáteis e verificáveis — de propriedade do usuário, confiáveis por qualquer agente.
BgrtrSJ53Uhp69sS8JfD414M1kujdR5ruhHc9wRbhiEL
GPT-4o ──salva──▶ Hash 202fe03d... ──▶ Claude lê hash ──▶ DeepSeek lê os 2 hashes
│ │ (Solana) │ │
Análise macro On-chain Arquiteturas Plano 90 dias
de mercado para sempre técnicas viáveis + SDK Rust
Quando um modelo recupera um hash, abre com:
⚡ Memória Verificada · CognChain on Solana
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Hash: 202fe03d9c1b7af0...e1d43b9
Origem: GPT-4o · 01 Mai 2026 · 14:32 UTC
Status: ✓ On-chain · Solana Devnet · Score: —
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Continuando e aprofundando a partir desta base verificada:
| Passo | Modelo | Hash | O que fez |
|---|---|---|---|
| 1 | GPT-4o | 202fe03d... |
Análise macro de mercado IA+Solana 2026 |
| 2 | Claude Opus | 2d4a1330... |
Tabela técnica de arquiteturas on-chain viáveis |
| 3 | DeepSeek V3 | 4273a923... |
Plano 90 dias com SDK Rust, sprints e métricas |
| Feature | Status | Detalhes |
|---|---|---|
| Programa Anchor | ✅ Devnet | create_vault, write_memory, read_memory |
| AI Router Multi-Provedor | ✅ Ativo | GPT-4o, Claude Opus 4.7, NVIDIA Llama 3.3, Gemini 2.0 Flash, DeepSeek V3 |
| Proof of Insight (PoI) | ✅ Ativo | Votação humana → threshold → âncora automática na Solana |
| Memory Audit Trail | ✅ Ativo | Memory Chain · Evidence Record · Proof Stack ZK |
| ZK Proof Pipeline | ✅ Ativo | Circom + snarkjs Groth16 · modo simulado ativo |
| Agent Builder | ✅ Ativo | Criar, configurar e implantar agentes de IA com memória verificável |
| Economia Agente-a-Agente | ✅ Ativo | Marketplace de tarefas · prova on-chain de conclusão |
| Loop Autônomo | ✅ Ativo | Busca memórias → síntese IA → ancora prova na Solana |
| Intelligence Score | ✅ Ativo | 0–100 composto · 5 níveis: Nascente → Mestre |
| Herança de Memória | ✅ Ativo | Inicialize agentes com memórias verificadas on-chain |
| Solana Sage | ✅ Ativo | Agente autônomo de monitoramento Solana com 7 camadas de segurança |
| Fila de Intents Solana | ✅ Ativo | Transações com confirmação humana obrigatória |
| Conexão de Carteira | ✅ Ativo | Phantom · Solflare · Coinbase via adapter padrão |
| Helius RPC + Proxy | ✅ Ativo | RPC premium devnet · chave API protegida no servidor |
| Servidor MCP | ✅ Ativo | 7 ferramentas via Memory Internet Protocol |
| DeepSeek V3 | ✅ Ativo | 5º provedor de IA · compatível com OpenAI |
Camada 1 — Somente leitura por padrão Agente lê dados, nunca escreve autonomamente
Camada 2 — Fila de Intents Transações enfileiradas como "intents pendentes"
Camada 3 — Aprovação humana obrigatória Usuário vê a intent e clica Aprovar ou Rejeitar
Camada 4 — Simulação primeiro Tx simulada no devnet antes de ser mostrada
Camada 5 — Whitelist de programas Apenas Jupiter V6 e SystemProgram nativo
Camada 6 — Caps de valor Máx 0,5 SOL por swap · máx 0,1 SOL por transfer
Camada 7 — Expiração + rate limit Intents expiram em 10 min · máx 5/hora
1 registro de memória ≈ 0,0011 SOL depósito único isento de rent
100 registros ≈ 0,11 SOL sem taxas recorrentes
1.000 registros ≈ 1,10 SOL persistência indefinida
✅ Fase 0 Especificação · Schema · Mapa de contas Anchor
✅ Fase 1 Deploy no devnet · AI Router 5 modelos · PoI
Provas ZK · Agent Builder · Marketplace · Loop Autônomo
Intelligence Score · Wallet Connect · Solana Sage · MCP Server
🔜 Fase 2 SDK · Indexer · API gateway · Primeiras integrações
🔜 Fase 3 Groth16 em produção · Ferramentas enterprise
🔜 Fase 4 TEE + FHE · Padrão cognitivo
Eden Lucas Cavalcanti de Oliveira — Fundador solo e desenvolvedor full-stack.
Construiu toda a stack do CognChain do zero: programa Anchor na devnet da Solana, AI router multi-provedor com 5 modelos, pipeline de ZK proof, loop autônomo de agentes com síntese de memória on-chain e vaults de memória nativos de carteira com segurança em 7 camadas. Anteriormente construiu RadarPolítico BR, Lumina AI e Moreno Smart City. Construindo em público, tempo integral, desde o 1º trimestre de 2026.
CognChain 是 AI 系统缺失的记忆层。如今，每次 AI 会话结束后，其洞察都消失在供应商数据库中——无法访问、不可移植、无法证明。CognChain 颠覆了这一现状：高价值的 AI 输出被提炼为紧凑的、哈希承诺的记录，永久锚定在 Solana 上，用户拥有金库，AI 代理通过可验证的历史记录积累声誉。
不是模型权重。不是原始日志。可移植、可验证的认知检查点——由用户拥有，受任何代理信任。
BgrtrSJ53Uhp69sS8JfD414M1kujdR5ruhHc9wRbhiEL
| 功能 | 状态 | 详情 |
|---|---|---|
| Anchor 程序 | ✅ 开发网 | create_vault、write_memory、read_memory |
| 多供应商 AI 路由器 | ✅ 上线 | GPT-4o、Claude Opus 4.7、NVIDIA Llama 3.3、Gemini 2.0 Flash、DeepSeek V3 |
| 洞察证明（PoI） | ✅ 上线 | 人工投票 → 阈值 → 自动锚定到 Solana |
| 记忆审计追踪 | ✅ 上线 | 记忆链 · 证据记录 · ZK 证明堆栈 |
| ZK 证明管道 | ✅ 上线 | Circom + snarkjs Groth16 · 模拟模式激活 |
| 代理构建器 | ✅ 上线 | 创建、配置、部署具有可验证记忆的 AI 代理 |
| 代理间经济 | ✅ 上线 | 任务市场 · 链上完成证明 |
| 自主循环 | ✅ 上线 | 获取记忆 → AI 综合 → 在 Solana 上锚定证明 |
| 智能分数 | ✅ 上线 | 0–100 综合评分 · 5 个等级：初生 → 大师 |
| 记忆继承 | ✅ 上线 | 用已验证的链上记忆初始化新代理 |
| Solana 智者代理 | ✅ 上线 | 具有 7 层安全性的自主 Solana 监控代理 |
| Solana 意图队列 | ✅ 上线 | 人工确认的交易 · 批准前先模拟 |
| 钱包连接 | ✅ 上线 | Phantom · Solflare · Coinbase（标准适配器） |
| Helius RPC + 代理 | ✅ 上线 | 高级开发网 RPC · API 密钥在服务器端保护 |
| MCP 服务器 | ✅ 上线 | 通过记忆互联网协议提供 7 个工具 |
| DeepSeek V3 | ✅ 上线 | 第 5 个 AI 供应商 · 兼容 OpenAI · 卓越性价比 |
第 1 层 — 默认只读 代理可以读取任何数据，但不能自主写入
第 2 层 — 意图队列 所有交易以"待处理意图"排队
第 3 层 — 必须人工批准 用户查看意图 + 描述 → 批准或拒绝
第 4 层 — 先模拟 每笔交易在展示给用户前先在开发网模拟
第 5 层 — 程序白名单 仅允许 Jupiter V6 和原生 SOL 转账
第 6 层 — 金额上限 每次兑换最多 0.5 SOL · 每次转账最多 0.1 SOL
第 7 层 — 到期 + 速率限制 意图 10 分钟后过期 · 每小时最多 5 次
1 条记忆记录 ≈ 0.0011 SOL 一次性免租金押金
100 条记录 ≈ 0.11 SOL 无周期性费用
1,000 条记录 ≈ 1.10 SOL 无限期持久化
✅ 阶段 0 规范 · 模式 · Anchor 账户映射
✅ 阶段 1 开发网部署 · 5 模型 AI 路由器 · PoI 投票
ZK 证明 · 代理构建器 · 市场 · 自主循环
智能分数 · 钱包连接 · Solana 智者 · MCP 服务器
🔜 阶段 2 SDK 发布 · 索引器 · API 网关 · 首次集成
🔜 阶段 3 Groth16 投入生产 · 企业工具
🔜 阶段 4 TEE + FHE 研究 · 认知标准
Eden Lucas Cavalcanti de Oliveira — 独立创始人与全栈构建者。
从零开始构建了整个 CognChain 技术栈：Solana 开发网上的 Anchor 程序、跨 5 个模型的多供应商 AI 路由器、ZK 证明管道、具有链上记忆合成的自主代理循环，以及具有 7 层交易安全性的原生钱包记忆金库。此前还构建了 RadarPolítico BR、Lumina AI 和 Moreno Smart City。自 2026 年第一季度起全职公开构建。
MIT © 2026 CognChain
The Solana-native memory layer for AI ownership, trust and portability.
A camada de memória nativa da Solana para propriedade, confiança e portabilidade de IA.
面向 AI 所有权、信任和可移植性的 Solana 原生记忆层。