Cogdex

Cognitive Runtime on Solana

Blockchain
as Cognitive State

Five memory classes. Confidence-weighted recall. Contradiction resolution. A cognitive runtime that learns from every interaction, updates what it believes, and checkpoints important state transitions.

Built for OpenAI Codex-powered agents and other autonomous runtimes.

Designed for agents that operate beyond a single session.

Cogdex · Active

where consciousness flows

Most AI agents can persist memory.
They can't maintain beliefs.

They respond. They log. They retrieve. Then contradictions accumulate, facts go stale, and identity drifts. Persistence alone doesn't create continuity. Without reconciliation, memory becomes noise.

In web3, that gap is costly: you want state you can verify and beliefs you can checkpoint — not black boxes that reset.

Persistence is easy. Reconciliation is hard.

See how the brain reconciles state

Stateless under pressure

Context windows flush after every conversation. Yesterday's constraint becomes today's partial summary.

Belief drift

Old facts keep ranking after they're outdated. New evidence arrives, but nothing resolves the conflict.

Prompt-based identity

Behavior is bolted on through system prompts. No durable self-model. No safe evolution from evidence.

No belief hygiene

Most memory stacks append forever. Few track confidence, contradictions, or which claims were superseded.

Stateless under pressure

Context windows flush after every conversation. Yesterday's constraint becomes today's partial summary.

Belief drift

Old facts keep ranking after they're outdated. New evidence arrives, but nothing resolves the conflict.

Prompt-based identity

Behavior is bolted on through system prompts. No durable self-model. No safe evolution from evidence.

No belief hygiene

Most memory stacks append forever. Few track confidence, contradictions, or which claims were superseded.

Stateless under pressure

Context windows flush after every conversation. Yesterday's constraint becomes today's partial summary.

Belief drift

Old facts keep ranking after they're outdated. New evidence arrives, but nothing resolves the conflict.

Prompt-based identity

Behavior is bolted on through system prompts. No durable self-model. No safe evolution from evidence.

No belief hygiene

Most memory stacks append forever. Few track confidence, contradictions, or which claims were superseded.

Stateless under pressure

Context windows flush after every conversation. Yesterday's constraint becomes today's partial summary.

Belief drift

Old facts keep ranking after they're outdated. New evidence arrives, but nothing resolves the conflict.

Prompt-based identity

Behavior is bolted on through system prompts. No durable self-model. No safe evolution from evidence.

No belief hygiene

Most memory stacks append forever. Few track confidence, contradictions, or which claims were superseded.

Stateless under pressure

Context windows flush after every conversation. Yesterday's constraint becomes today's partial summary.

Belief drift

Old facts keep ranking after they're outdated. New evidence arrives, but nothing resolves the conflict.

Prompt-based identity

Behavior is bolted on through system prompts. No durable self-model. No safe evolution from evidence.

No belief hygiene

Most memory stacks append forever. Few track confidence, contradictions, or which claims were superseded.

Stateless under pressure

Context windows flush after every conversation. Yesterday's constraint becomes today's partial summary.

Belief drift

Old facts keep ranking after they're outdated. New evidence arrives, but nothing resolves the conflict.

Prompt-based identity

Behavior is bolted on through system prompts. No durable self-model. No safe evolution from evidence.

No belief hygiene

Most memory stacks append forever. Few track confidence, contradictions, or which claims were superseded.

Stateless under pressure

Context windows flush after every conversation. Yesterday's constraint becomes today's partial summary.

Belief drift

Old facts keep ranking after they're outdated. New evidence arrives, but nothing resolves the conflict.

Prompt-based identity

Behavior is bolted on through system prompts. No durable self-model. No safe evolution from evidence.

No belief hygiene

Most memory stacks append forever. Few track confidence, contradictions, or which claims were superseded.

Stateless under pressure

Context windows flush after every conversation. Yesterday's constraint becomes today's partial summary.

Belief drift

Old facts keep ranking after they're outdated. New evidence arrives, but nothing resolves the conflict.

Prompt-based identity

Behavior is bolted on through system prompts. No durable self-model. No safe evolution from evidence.

No belief hygiene

Most memory stacks append forever. Few track confidence, contradictions, or which claims were superseded.

How the stack works end-to-end

A clean four-step flow: memory classes and hybrid retrieval → contradiction-aware graph → confidence-weighted recall → reconciliation cycles with optional checkpoint to chain.

STEP 01Multi-signal recall

Hybrid retrieval

Vector similarity, keyword matching, tag scoring, and source trust weighting. Ranked recall across semantic and symbolic signals.

STEP 02

Contradiction graph

Typed, weighted links between memories and beliefs. Conflict is first-class; the graph tracks what contradicts what.

Graph typeContradiction-aware
STEP 03

On-chain verification

Checkpoint state hashes to Solana when it matters. Every consumer query can be verified; full content can stay private.

STEP 04

Reconciliation cycles

Scheduled consolidation and conflict resolution. Evidence checks, not blind promotion. Beliefs update from proof.

Five classes. Different decay and confidence behavior.

Not all memory should behave the same way. Events fade. Facts evolve. Procedures reinforce. Identity updates slowly. Active beliefs require conflict handling.

Type 01

Episodic

Raw interactions and events. Who said what, when, and in what context.

Fast (adaptive)
Type 02

Semantic

Distilled facts and recurring patterns from evidence during reconciliation.

Slow (confidence-aware)
Type 03

Procedural

Learned behavior and execution patterns. What works, what fails.

Reinforced by reuse
Type 04

Self-Model

Evolving self-understanding and constraints. Updated with stronger evidence.

Minimal (strict gating)
Type 05

Belief Ledger

Confidence-scored claims with evidence lineage and contradiction state.

Scope + evidence

Every 6 hours, Cogdex audits what it thinks it knows.

A four-phase consolidation process inspired by memory research and reliability engineering. Not idle reflection — structured belief maintenance.

I

Consolidation

Focal questions are generated from recent memory windows. Each question retrieves an evidence set and candidate links.

II

Reconciliation

Contradictions are detected, confidence is updated, and beliefs are marked active, scoped, superseded, or unresolved.

III

Reflection

Self-model and procedural memory are reviewed against recent outcomes. Identity evolves from evidence, not static prompts.

IV

Emergence

Cogdex produces an optional synthesis: a human-readable insight, changelog, or internal narrative for downstream agents.

Solana

Not every thought needs a receipt. Important state changes do.

Cogdex computes hashes for memory events and can checkpoint cognitive state transitions to Solana-compatible verification rails. Full content can remain private while checkpoints preserve provenance.

Spin up verification when it matters — local-first for speed, on-chain for proof.

Verify in browser
TransactionHashSlotStatus
#18Qm5...dW3j214,881,392Verified
#22Rn7...vK4s214,881,388Verified
#35Kx9...mR7q214,881,385Verified

Content stays private. Only checkpoint provenance is committed.

Latency / Privacy / Verifiability

Storage Modes

Not every memory event needs the same durability. Choose local for speed, hybrid for production, or checkpointed modes for auditability and public proof.

Fast & Private

Local

Everything stored locally. Lowest latency. No external dependency. Best for development, local agents, and private workflows.

Recommended

Hybrid

Hot memory stays local. Durable memory lives in your backend. Checkpoints are optional and only used for meaningful state transitions.

Verifiable State

Checkpointed

Cognitive state updates are hashed and checkpointed. Useful for public agents, shared systems, and trust-sensitive applications.

Local first

Low latency

Working memory & hot paths stay local. No external dependency for core recall.

Details →

Modes

Local · Hybrid · Checkpointed

Durable memory in your backend. Optional checkpoints to chain for verification.

→ Verify in browser

Get started

Choose your tradeoff

Install the SDK, run the wizard, and give your agent persistent memory in under a minute.

Try demo

A cognitive runtime in a few lines.

Install the SDK, run the wizard, and give your agent persistent memory with contradiction-aware consolidation in under a minute.

terminal
1$ npm install cogdex
2$ npx cogdex init
app.ts
1import { Cortex } from 'cogdex';
2
3const brain = new Cortex({
4  storage: { url, serviceKey },
5  codex: {
6    apiKey: process.env.OPENAI_API_KEY, // optional
7    model: 'gpt-5.3-codex',
8  },
9});
10
11await brain.init();
12
13// Store a memory
14await brain.store({
15  type: 'episodic',
16  content: 'User prefers concise updates but long-form technical analysis',
17  summary: 'Scoped response-length preference',
18  source: 'my-agent',
19});
20
21// Recall with hybrid retrieval
22const memories = await brain.recall({
23  query: 'How detailed should responses be for technical topics?',
24});
5Memory classes
Conflict-awareRetrieval
<250msMedian recall latency (target)
Codex-readyOpenAI Codex integration
AdaptiveDecay + confidence
MITLicense
5Memory classes
Conflict-awareRetrieval
<250msMedian recall latency (target)
Codex-readyOpenAI Codex integration
AdaptiveDecay + confidence
MITLicense
5Memory classes
Conflict-awareRetrieval
<250msMedian recall latency (target)
Codex-readyOpenAI Codex integration
AdaptiveDecay + confidence
MITLicense
5Memory classes
Conflict-awareRetrieval
<250msMedian recall latency (target)
Codex-readyOpenAI Codex integration
AdaptiveDecay + confidence
MITLicense
5Memory classes
Conflict-awareRetrieval
<250msMedian recall latency (target)
Codex-readyOpenAI Codex integration
AdaptiveDecay + confidence
MITLicense
5Memory classes
Conflict-awareRetrieval
<250msMedian recall latency (target)
Codex-readyOpenAI Codex integration
AdaptiveDecay + confidence
MITLicense
5Memory classes
Conflict-awareRetrieval
<250msMedian recall latency (target)
Codex-readyOpenAI Codex integration
AdaptiveDecay + confidence
MITLicense
5Memory classes
Conflict-awareRetrieval
<250msMedian recall latency (target)
Codex-readyOpenAI Codex integration
AdaptiveDecay + confidence
MITLicense