oceanbase/powermem

What it solves

AI agents often struggle with finite context windows and noisy long-term memory. PowerMem provides a persistent, self-evolving memory system that transforms raw conversations and actions into structured, searchable long-term memory, preventing the "static note dump" problem common in naive save-everything approaches.

How it works

PowerMem uses an LLM-driven pipeline to extract, update, and merge memories. It employs a hybrid retrieval system combining vector, full-text, and graph search with recency signals. To keep memory relevant, it implements Ebbinghaus-style time decay (forgetting curves). A unique "Experience + Skill distillation" process allows agents to distill reusable workflows from interactions, enabling the system to evolve over time.

Who it’s for

Developers building long-running AI agents, copilots, or multi-agent systems that need to maintain user profiles, project decisions, and learned workflows across multiple sessions.

Highlights

• Self-Evolving Memory: Two-layer distillation of experiences and skills to learn reusable workflows.
• Hybrid Retrieval: Out-of-the-box integration of vector, full-text, and graph retrieval.
• Intelligent Lifecycle: LLM-driven extraction and Ebbinghaus-style decay to manage memory noise.
• Broad Integration: Supports Python SDK, HTTP server, MCP (Model Context Protocol), and plugins for IDEs like Cursor and VS Code.
• Multimodal Support: Handles text, image, and audio signals.