PowerMem is a persistent, self-evolving memory layer for AI agents and applications, developed in the OceanBase ecosystem. It addresses the failure mode of naive "save everything" memory — noisy, expensive, hard to retrieve — by turning conversations, actions, and feedback into structured long-term memory with an LLM-driven lifecycle: extraction, update, merge, and Ebbinghaus-style time decay. Its distinguishing feature is two-layer Experience + Skill distillation, which lets agents learn reusable workflows from interactions rather than only recalling facts. It also supports multi-agent isolation and sharing, user profiles, and multimodal signals (text, image, audio).
Retrieval is hybrid out of the box, combining vector, full-text, graph, and recency signals in one layer. One backend serves several integration surfaces — Python SDK (with Go, Java, and TypeScript SDKs in sibling repos), an HTTP REST server with a web dashboard, an MCP server, and a pmem CLI — plus first-party plugins and setup guides for Claude Code, Cursor/VS Code/Windsurf/Copilot, OpenClaw (via a memory plugin that defaults to a local SQLite-backed CLI mode), LangChain/LangGraph, AgentScope, and any generic MCP client. Storage backends include OceanBase (with graph support), an embedded seekdb for zero-config local use, PostgreSQL/pgvector, and SQLite; a wide matrix of LLM, embedding, and rerank providers is built in.
Reported benchmark results (reproducible from the repo's benchmark directory): 87.79% accuracy on LOCOMO versus a 52.9% baseline, with p95 search latency of 1.44s and roughly 0.9K tokens per query, and on AppWorld a 39% pass rate versus 24% baseline with fewer steps and about a third fewer tokens.
