AgentDB

•
109
•
已认领
管理

Provides a self-learning vector memory that improves search quality with agent feedback, consolidating vectors, indexes, and learning state into a single file.

AgentDB offers a unique approach to vector memory, moving beyond simple embedding storage to actively learn from agent interactions. It observes which search results an agent utilizes, then leverages this feedback through a Thompson Sampling bandit to enhance subsequent queries, optimizing ranking, reinforcement learning algorithm selection, and compression tiers autonomously. This cognitive container integrates episodic memory, skill libraries, causal reasoning, and a learning loop, all encapsulated in a portable `.rvf` file, making the database itself smarter over time and running efficiently across various environments like Node, browser, and edge runtimes.

主要功能

0155 GitHub stars
02Self-Learning Search with up to +36% quality lift from agent feedback
03High-performance RuVector Rust engine offering sub-millisecond reads on 100k+ vectors
04Single-file '.rvf' Cognitive Container for all data, indexes, and learning state
05Six human-inspired cognitive memory patterns including episodic replay and skill libraries
0641 MCP tools for seamless integration with LLM agents like Claude Code and Cursor

使用案例

01Integrating intelligent, self-learning vector memory into custom Node.js, browser, or edge applications
02Enhancing LLM agents with advanced cognitive patterns, causal reasoning, and feedback-driven memory via MCP tools
03Developing AI applications requiring fast, efficient, and offline-capable memory with reinforcement learning capabilities