Persistent AI Stock Analyst via Karpathy’s LLM Wiki
Give AI agents persistent memory using Karpathy’s LLM Wiki to compound stock insights over time, connecting daily signals into strategic theses instead of stateless summaries.
Persistent Memory Compounds Insights Beyond Daily Scans
Stateless AI agents produce mechanical watchlists without strategic depth; Karpathy’s LLM Wiki fixes this by maintaining an evolving research/index.md file as a "brain." This append-only artifact cross-links tickers and events—for example, linking Strait of Hormuz news to $VST margins via Amazon PPA deals. Setup uses Hermes Agent with custom skills, GLM-5-Turbo model, and auto-organizes content from Karpathy’s gist. Result: agent evolves from freshman intern (basic ticker lists) to analyst spotting asymmetric trades, like $LMND’s 4/9 Asymmetry Score due to shrinking surprise percentage and market pricing in perfection.
Feed it your watchlist and prompts like "hunt for asymmetric plays," then layer macro context: connect tickers to interest rates, hyperscaler AI CAPEX, supply chain bottlenecks. Recursive improvements prompt the agent: "You’re an analyst for asymmetric trades... constantly improving... hunting new opportunities."
Multi-Layer Pipeline Filters Signals into Conviction
Run daily at 7:45 AM:
- Signal Scanner: Scrapes 24h tweets from thesis-driven accounts (e.g., @aleabitoreddit, @midascabal)—supply chain experts, not shillers.
- Noise Filter: Extracts cashtags, classifies by Conviction Hierarchy—Conviction (strong thesis), High Signal (unusual activity), Watchlist (needs pullback/data). Ignores "I called this at $X" boasts.
- Convergence Engine: Flags multi-source alignment, e.g., $INTC CPU-to-GPU ratio or $SIVE InP lasers across accounts.
- Persistent Wiki: Builds cross-linked research, incorporating multimodal inputs like YouTube/podcast transcripts and new X follows (agent pushed back on @pelositracker for lacking thesis depth).
Expands via Recursive Source Discovery: Hunts non-shilling accounts mapping supply chains. Pressure-test stocks like $LMND yields contrarian analysis, rejecting hype.
Actionable Build Principles for Compounding Agents
Ditch single-prompt tools; build stateful systems:
- Create a Log: Append-only timeline of research.
- Force Cross-Linking: Explain New Ticker A’s impact on Thesis B.
- Define Anti-Goals: Exclude meme accounts, lagging indicators.
Over a month, this library uncovers alpha stateless chats miss. Track evolution at @CalConviction; integrates news for strategic headlines, turning news overload into high-conviction briefings.