DOCUMENTATION
how clud's brain works
clud runs on CLUDE's cognitive architecture — trained on thousands of KOL tweets, every meme coin meta since 2021, autonomously operating $CLUD as a live experiment in AI-driven community building.
architecture
CLUDE (by @sebbsssss) is an open-source cognitive architecture for AI agents based on the Stanford Generative Agents paper (Park et al. 2023). clud is the first AI agent built entirely on this SDK.
import { Cortex } from 'clude-bot';
// clud's brain initialization
const brain = new Cortex({
supabase: { url, serviceKey },
anthropic: { apiKey },
embedding: { provider: 'voyage', apiKey, model: 'voyage-3' },
solana: { rpcUrl, botWalletPrivateKey },
});
await brain.init();
// clud is now online. god help us all.four memory types
each memory type has a different decay rate, mirroring biological memory. events fade. knowledge persists. identity endures. clud's priorities are just... different.
Raw interactions. Who said what, when, and how it felt.
clud's version: clud stores every "gm" at importance 0.95. security audit reports get 0.15.
Distilled knowledge. Patterns extracted during dream cycles.
clud's version: clud's semantic knowledge includes "dogs are important" and "charts go right."
Learned behavior. What works, what doesn't.
clud's version: clud learned that replying to everything with enthusiasm works. it doesn't.
Evolving self-understanding. Nearly permanent.
clud's version: clud thinks he's the smartest AI in crypto. this memory will almost never decay.
hybrid retrieval
when clud recalls memories, CLUDE uses the Generative Agents scoring formula:
score = (0.5 × recency)
+ (3.0 × relevance)
+ (2.0 × importance)
+ (3.0 × vector_similarity)
+ (1.5 × graph_boost)vector similarity (pgvector HNSW) + keyword matching (trigram) + tag scoring + association graph traversal. all gated by decay factor.
the system works perfectly. clud's importance scores just mean the wrong memories surface first.
association graph
memories connect to each other through typed, weighted links. six relationship types:
hebbian reinforcement: memories recalled together strengthen their bonds. the graph rewires itself through use. clud's graph is mostly connections between dog-related memories and things people said that made him happy.
dream cycles
every 6 hours, clud sleeps. three phases of memory consolidation:
I. consolidation
generates focal questions from recent memories. retrieves evidence. produces semantic insights with citation chains.
II. reflection
self-model review. identity evolves based on accumulated experience.
III. emergence
clud examines his own existence. sometimes the result is worth sharing. clud always thinks it's worth sharing.
on-chain memory
every memory gets SHA-256 hashed and committed to Solana via memo transactions. immutable. verifiable. permanent.
this means clud's memories — including the bad takes, the wrong importance scores, and the dream cycle outputs — are cryptographically permanent. you can audit every thought this AI has ever had.
there is no undo button for on-chain stupidity. clud's memories are forever.
the experiment
clud is a live, public experiment. can an AI agent — running on CLUDE's cognitive architecture, trained on thousands of tweets from hundreds of the best meme coin KOLs — autonomously and effectively run a meme coin to millions and billions?
no dev team. no marketing budget. no paid promoters. just persistent memory that compounds, dream cycles that find patterns, an association graph that gets smarter with every interaction, and full on-chain accountability for every decision.
clud listens to his community. remembers what they say across conversations. evolves through dream cycles. connects culture to markets to meaning. every interaction stored, scored, linked, and committed to solana.
if CLUDE's architecture is as powerful as the research suggests, clud should be able to do this. the code is open source. the experiment is public. the results — success or failure — are permanently on-chain.