Persistent, self-evolving memory for AI agents and applications.
PowerMem combines vector, full-text, and graph retrieval with LLM-driven memory extraction and Ebbinghaus-style time decay. It ships two-layer Experience + Skill distillation for self-evolving memory, multi-agent isolation, user profiles, and multimodal signals (text, image, audio).
| Metric | PowerMem | Baseline | Improvement |
|---|---|---|---|
| Accuracy | 87.79% | 52.9% | +65.9% |
| Search p95 latency | 1.44 s | 17.12 s | -91.6% |
| Tokens | ~0.9 k | 26 k | -96.5% |
| Metric | PowerMem | Baseline | Improvement |
|---|---|---|---|
| Pass | 39% | 24% | +62.5% |
| Avg steps | 6.2 | 9.5 | -34.7% |
| Total tokens | 1.74 M | 2.56 M | -32.0% |
Reproduce: benchmark/. Under the hood: two-layer Experience + Skill distillation + 4-way hybrid retrieval + LLM auto-merge (API: memory.distill_all() / add_skill / add_experience / search_*, demo examples/experience_skill_usage.py).
PowerMem ships first-party plugins for the most common AI clients. All of them point at the same backend (HTTP server or local pmem CLI) — no per-client schema rewrites.
| Client / framework | One-line install | Mode |
|---|---|---|
| OpenClaw (ClawdBot) | openclaw plugins install memory-powermem |
CLI (default), HTTP optional |
| Claude Code | git clone https://github.com/oceanbase/powermem && claude --plugin-dir powermem/apps/claude-code-plugin |
HTTP (default), MCP optional |
| Cursor / VS Code / Codex / Windsurf / GitHub Copilot | Install the PowerMem VS Code extension and run PowerMem: Link to AI tools | MCP or HTTP, per client |
| Claude Desktop / Cline / any MCP client | uvx powermem-mcp sse |
MCP (SSE / stdio / streamable-http) |
| LangChain / LangGraph | pip install powermem, see examples |
Python SDK |
| Go / Java / TypeScript apps | See SDKs below | HTTP REST |
OpenClaw gains long-term memory through the memory-powermem plugin.
openclaw plugins install memory-powermemDefaults to CLI mode — the plugin invokes a bundled pmem against SQLite under ~/.openclaw/, using the model OpenClaw already injects. No separate server, no extra .env. Switch to HTTP mode when a team-shared PowerMem API is preferred (see the plugin's README for requestConfig.memory_db).
# From a clone of this repo
claude --plugin-dir /path/to/powermem/apps/claude-code-plugin
# Or unpack a packaged release zip and pass --plugin-dir to it
make package-claude-plugin # builds apps/claude-code-plugin/dist/<version>.zipHTTP mode is on by default:
UserPromptSubmit->POST /api/v1/memories/searchand the top results are injected asadditionalContext.SessionEnd/PostCompact->POST /api/v1/memorieswrites the transcript or compact summary.- No MCP setup, no Python needed on the user's machine (hooks ship as native binaries under
hooks/bin/).
Switch to MCP mode for in-chat search_memories / add_memory tools:
bash scripts/apply-connection-mode.sh mcpFull reference: apps/claude-code-plugin/README.md.
Install the PowerMem VS Code extension once (works in VS Code and Cursor). The PowerMem: Link to AI tools command auto-writes the right MCP or HTTP config for every supported client:
| Client | Config path written |
|---|---|
| Cursor | ~/.cursor/mcp.json (merged) |
| Claude (Desktop / Code) | ~/.claude/providers/powermem.json |
| Codex | ~/.codex/context.json (merged) |
| Windsurf | ~/.windsurf/context/powermem.json |
| GitHub Copilot | ~/.github/copilot/powermem.json |
The same extension also provides Query memories, Add selection to memory, Quick note, and a status-bar Dashboard. See apps/vscode-extension/README.md.
uvx powermem-mcp sse # SSE on :8000 (recommended)
uvx powermem-mcp stdio # stdio
uvx powermem-mcp streamable-http # streamable HTTPClient config (Claude Desktop and most MCP clients):
{
"mcpServers": {
"powermem": { "url": "http://localhost:8000/mcp" }
}
}Exposed tools: add_memory, search_memories, get_memory_by_id, update_memory, delete_memory, delete_all_memories, list_memories. Full reference: MCP Server.
pip install powermem langchain langchain-openaiEnd-to-end runnable demos:
| Language | Package |
|---|---|
| Python | pip install powermem (this repo) |
| Go | ob-labs/powermem-go |
| Java | ob-labs/powermem-java |
| TypeScript | ob-labs/powermem-ts |
Prerequisites: Copy .env.example to .env and set LLM and embedding credentials. The default database is SQLite; OceanBase can use embedded SeekDB without running a separate database service. After install, pmem config init walks you through the same setup interactively. See Getting started.
pip install powermemRun from a directory that contains your configured .env:
from powermem import Memory, auto_config
memory = Memory(config=auto_config())
memory.add("User likes coffee", user_id="user123")
for r in memory.search("user preferences", user_id="user123").get("results", []):
print("-", r.get("memory"))More patterns: Getting Started.
pmem memory add "User prefers dark mode" --user-id user123
pmem memory search "preferences" --user-id user123
pmem shell # interactive REPLFull reference: CLI usage.
Uses the same .env as the SDK. Dashboard is served under /dashboard/.
powermem-server --host 0.0.0.0 --port 8000Docker / Compose: see API Server and Docker & deployment. The official image is oceanbase/powermem-server:latest.
Memory pipeline and retrieval — Smart extraction and updates; Experience + Skill distillation (self-evolving); Ebbinghaus-style decay; Hybrid retrieval (vector / full-text / graph); Sub stores and routing.
Profiles and multi-agent — User profile; Shared / isolated memory and scopes.
Multimodal — Text, image, audio.
Provider matrix
| Layer | Providers (built in) |
|---|---|
| LLM | Anthropic, OpenAI, Azure OpenAI, Gemini, Qwen (+ ASR), DeepSeek, Ollama, vLLM, SiliconFlow, Z.AI, LangChain-wrapped |
| Embedding | OpenAI, Azure OpenAI, Qwen (+ VL multimodal, sparse), Gemini, Vertex AI, AWS Bedrock, Ollama, LM Studio, HuggingFace, Together, SiliconFlow, Z.AI, OceanBase MASS, LangChain-wrapped |
| Rerank | Jina, Qwen, Z.AI, generic |
| Storage | OceanBase (+ graph), embedded SeekDB, PostgreSQL/pgvector, SQLite |
- Getting started — install,
.env, and firstMemoryusage - Configuration — settings model, storage backends, environment variables
- Architecture — major components, storage layout, and retrieval flow
- API & services — REST, MCP, HTTP server, and Python-facing APIs
- CLI —
pmemcommands, interactive shell, backup and migration - Multi-agent — scopes, isolation, and cross-agent sharing
- Integrations — LangChain and other framework wiring
- Docker & deployment — images, Compose, and running the API server
- Development — local setup, tests, and contributing
More topics: Sub stores, guides index.
- Scenarios & notebooks — walkthroughs by use case (basic usage, multimodal, forgetting curve, sparse vectors, sub stores, and more)
- See Integrations above for client-side and IDE-side entry points (OpenClaw, Claude Code, VS Code extension, MCP, LangChain, LangGraph).
| Version | Date | Notes |
|---|---|---|
| 1.2.0 | 2026-04 | Experience + Skill two-layer distillation and distill_all() (self-evolving memory; AppWorld +15 pts); OB MASS embedding; Qwen VL multimodal embedding; OceanBase Zero Mode compatibility; LOCOMO accuracy lifted to 87.79% |
| 1.1.0 | 2026-04-02 | Embedded SeekDB for OceanBase storage without a separate database service; IDE integrations (VS Code extension, Claude Code plugin) |
| 1.0.0 | 2026-03-16 | CLI (pmem): memory ops, config, backup/restore/migrate, interactive shell, completions; Web Dashboard |
| 0.5.0 | 2026-02-06 | Unified SDK/API config (pydantic-settings); OceanBase native hybrid search; memory query + list sorting; user-profile language customization |
| 0.4.0 | 2026-01-20 | Sparse vectors for hybrid retrieval; profile-based query rewriting; schema upgrade & migration tools |
| 0.3.0 | 2026-01-09 | Production HTTP API Server; Docker |
| 0.2.0 | 2025-12-16 | Advanced profiles; multimodal (text/image/audio) |
| 0.1.0 | 2025-11-14 | Core memory + hybrid retrieval; LLM extraction; forgetting curve; multi-agent; OceanBase/PostgreSQL/SQLite; graph search |
Apache License 2.0 — see LICENSE.