This week, a new project climbed the GitHub trending charts with a provocative premise: instead of paying broker commissions and AI subscription fees for automated trading, developers should build their own investment agent from open-source code. The repository drops the entire stack — account management, trade execution, memory systems, and tool integrations — into the public domain.

20 Tools and a Three-Tier Memory System, All in One Agent

The architecture the team published is deliberately transparent. Account management uses event sourcing, recording every state change as an immutable event so all trade histories can be replayed and audited. Before any batch execution, the system runs an event-driven validation pass; only after that check does the agent place real orders. Rather than orchestrating multiple specialized agents, a single agent handles everything — it calls tools, consults memory, and makes decisions alone. But before finalizing a trade, multiple agent nodes share their reasoning in a verification step, and after execution, the agent posts entertaining forum-style commentary for the user.

The tool inventory includes roughly 20 items beyond standard API calls. One notable custom tool stacks four ML models in BigQuery, combining their predictions to improve accuracy. The full tool list is available in the repository.

The memory system is where the project shows its most deliberate design. It pulls from both a vector database for similarity search and raw BigQuery data. Memory is organized into three tiers based on importance, with an exponential forgetting curve that decays weights over time. Memories are linked in a graph structure, so retrieving one pulls in related context. Only memories validated through an ontology — a structured knowledge graph of entity relationships — survive as semantic memory.

Here is how it works in practice. If the agent buys Apple (AAPL) right before earnings while the RSI signals overbought, and the trade loses money, that loss gets high importance and lands in episodic (medium-term) memory. If a similar pattern repeats, the lesson graduates to semantic (long-term) memory. Routine trades lose weight automatically after two weeks, but large gains or losses, and frequently queried memories, decay slowly. The next time the agent evaluates Apple, it does not just recall the loss — it pulls in related memories like "China supply chain concerns" and "semiconductor sector weakness." The ontology filters out vague impressions like "Apple feels risky" and keeps only structured relationships such as "Apple --risk--> China supply chain."

From Black Box to Open Hood: Every Line of Code Is Editable

Traditional AI investment services are black boxes. Users pay fees, submit capital, and receive trade signals with no visibility into the reasoning. This project inverts that model. The UI exposes nearly everything: the tool codebase, every prompt injected into the LLM, and the raw outputs of tool executions. Customization covers model settings, risk management parameters, investment strategy prompts, forum post style, tool selection, and the entire three-tier memory system.

If the provided tools do not fit a user's strategy, they can build their own and attach them via the Model Context Protocol (MCP), the standard interface for agent-to-tool communication. The repository also includes two supplementary agents: one for news retrieval and another that summarizes the investment agent's post-batch behavior and stores it in memory.

Performance is still early-stage. In the U.S. market, the agent has underperformed the market benchmark. In the domestic Korean market, it has roughly matched the index. The development team acknowledges this openly, framing the project's value proposition differently: "Investing will eventually shift to agents. The question is whether you pay fees for someone else's agent or build your own and keep the fees." By releasing the code as open source, they argue, developers can avoid paying both brokerage commissions and agent subscription fees, iterating on their own strategies instead.

Whether the investment agent market moves from closed, fee-based services to open-source self-building depends on whether the performance numbers eventually back up the philosophy. For now, the code is on the table.