Agentic AI Evolves With Self-Improving Memory
Agentic AI systems for quantitative finance are advancing from static tools to self-improving engines with persistent, multi-tiered memory. A "QuantAgent" workflow reportedly uses a continuous feedback loop to track, score, and refine alpha signals based on live performance. This architecture, which requires robust state management, is being productionized with tiered memory systems using low-latency session stores and permanent graph databases to solve the statelessness that limits most AI agents, according to Cognee founder Vasilije Markovich on the Data Engineering Podcast.
- The tiered memory architecture addresses the high cost and performance degradation that occurs when large, undifferentiated memory files are fed into an LLM's context window. This "context rot" is a key failure point for agents that need to operate over long periods. Systems like MemGPT manage this by treating the context window like RAM and external databases like disk storage, allowing the agent to page memory in and out as needed. - Cognee, the company founded by Vasilije Markovich, recently raised $7.5 million in a seed round led by Pebblebed to build a structured, persistent memory layer for AI agents. The company aims to solve the problem of AI "amnesia" by transforming unstructured data into a knowledge graph, moving beyond simple text retrieval to understand context and relationships within the data. - Production-grade agentic systems often require a hybrid approach to memory, combining different database types for specific functions. This includes using low-latency in-memory stores like Redis for session state and caching, vector databases for semantic search of past interactions, and relational databases for structured state management and procedural memory. - The use of graph databases is critical for representing complex relationships in financial data, which is a limitation of traditional relational databases. In agentic systems, knowledge graphs allow the AI to discover non-obvious connections between assets or entities, moving beyond simple data retrieval to a more deterministic understanding of market dynamics. - This shift towards stateful, agentic AI is creating new architectural patterns in quantitative finance, moving beyond model improvements to focus on the systems that manage data and workflow orchestration. The goal is to create autonomous systems that can discover signals, evolve strategies, and adapt execution in dynamic markets. - Vasilije Markovich's work on Cognee is influenced by his background in cognitive science, specifically the multilayered models of human language and memory. This approach structures AI memory into different layers, similar to how humans store individual words, phrases, and sentences, to enable more efficient retrieval.
