I built HashTrade — an open-source autonomous trading agent that treats an LLM as a non-parametric decision function conditioned on episodic memory, rather than encoding strategy as code.

The core idea: instead of writing if/else trading logic, you give an LLM three tools (exchange access, memory, UI control) and let it form strategy through accumulated experience. The agent wakes on a variable 5→10→20→25 min cycle, reads its past notes, fetches market data, reasons about what to do, and optionally executes trades. Every decision and outcome is logged to an append-only JSONL file that becomes its long-term memory.

Technical details:

- Built on Strands Agents (AWS) with CCXT for 100+ exchange support - 3 tools only: use_ccxt (28 actions — market data, orders, arbitrage detection), history (persistent memory), interface (dynamic UI) - Variable-interval scheduler to avoid detectable timing patterns in order flow - Fire-and-forget WebSocket streaming for sub-second dashboard latency - Supports Claude, GPT-4o, Ollama (local), and Bedrock — auto-detected - PWA frontend in vanilla JS, no framework dependencies - Client-side credential isolation — API keys never leave the browser - Recursive credential redaction prevents keys from leaking into LLM context

The interesting emergent behavior: early wake cycles are conservative ("observing BTC at $67k, noting support level"). After a few days of accumulated memory, the agent starts referencing its own past observations to form trading theses ("last 3 times we saw this pattern, price bounced — going long"). The policy improves not through fine-tuning but through growing context.

I wrote a paper formalizing this as a Memory-Conditioned Markov Decision Process if anyone's interested in the theory: the key insight is that the effective policy is non-stationary even with fixed model parameters, because the growing memory changes the attention distribution at each step.

Setup: pip install hashtrade && hashtrade

Live demo at hashtrade.ai, code at github.com/mertozbas/hashtrade. Apache 2.0.

Would love feedback on the architecture — especially the tradeoff between soft risk constraints (enforced via system prompt) vs. hard tool-level enforcement.

> This isn't another trading bot. It's an autonomous agent with memory, intuition, and the authority to act.

The AI smell sure is strong with this one.

Overall, this bot will fail to generate a profit; it is trading what is already confirmed, using data that everyone has, ergo there is no edge.