Show HN: Agent Memory DB – Episodic memory for agents (97.5% vs. 85% recall) (github.com)

A new tool called Agent Memory DB has been introduced, providing AI agents with episodic memory capabilities. This innovative database allows agents to store experiences and retrieve relevant past encounters based on similarity instead of simply relying on recent interactions. The tool significantly boosts recall success rates, achieving an impressive 97.5% in a 40-task evaluation, compared to a mere 85% when using truncated history methods. This improvement is particularly notable in long-history tasks, where recall jumped from 33.3% to 88.9%. The significance of Agent Memory DB lies in addressing the limitations of large language models (LLMs), which have constrained context windows that often truncate essential information. By enabling agents to maintain a comprehensive and retrievable memory, developers can enhance the performance of AI applications across various tasks. Technically, the memory system supports multiple programming languages, including Rust, Python, Node.js, and Go, leveraging fast approximate nearest neighbor (HNSW) vector search for quick retrieval. Furthermore, with a query latency of around 200 microseconds and the capability to handle over 10,000 inserts per second, Agent Memory DB is designed for efficient use in dynamic environments.