PowerMem is a persistent memory layer for AI agents and applications, built by OceanBase. Its successor, PowerContext, keeps the context of a piece of work across conversations, so that a person or a new agent can take over a task and continue from its current state without rereading the full history.
PowerContext stores durable facts as Memory, organizes the current objective and state into a Handoff, and records reusable approaches as Experience or Skill. Every item stays within the scope of the work it belongs to and keeps its sources and earlier revisions. With full memory capabilities enabled, it turns prompts into Sources, groups them into Topics that evolve with related prompts, and recalls them in later sessions within the same Scope. A basic mode saves and retrieves memories explicitly without extra model APIs.
The memory engine extracts key facts from conversations with an LLM, detects duplicates, updates conflicting information, and merges related memories, with time-decay weighting that favors recent ones. Retrieval combines vector search, full-text search, and graph retrieval, and images and audio are converted to text for storage. Each agent gets its own memory space, with scoped sharing across agents.
PowerContext runs as a server that agents reach through plugins for Codex, Claude Code, Hermes Agent, OpenClaw, OpenCode, Pi Coding Agent, and other hosts, and through integrations with Pydantic AI, LangChain, and LangGraph. Applications use it through an async Python client, an HTTP API, MCP, or an in-process Core SDK.
Features
- Handoff: organizes a task's objective, decisions, constraints, progress, evidence, and next steps so another person or agent can continue it
- Memory, Experience, and Skill: separate stores for durable facts, reusable approaches, and recorded skills, each kept with its sources and revisions
- Automatic memory: turns prompts into Sources and Topics that evolve over time and are recalled in new sessions
- LLM fact extraction: extracts key facts, detects duplicates, resolves conflicting information, and merges related memories
- Time decay: weights recent and relevant memories higher during recall
- Hybrid retrieval: combines vector, full-text, and graph retrieval, with sub-stores and automatic query routing
- Multimodal memory: converts images and audio to text descriptions and retrieves mixed content
- Multi-agent scopes: a memory space per agent, with cross-agent sharing and scope-based permissions
- Agent integrations: plugins for Codex, Claude Code, Hermes Agent, OpenClaw, OpenCode, and Pi Coding Agent, plus Pydantic AI, LangChain, and LangGraph
- Interfaces: async Python client, HTTP API, MCP server, and an in-process Core SDK
