powermem 1.1.2

Intelligent Memory System - Persistent memory layer for LLM applications

PowerMem

Persistent, self-evolving memory for AI agents and applications.

English · 中文 · 日本語

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).

---

Benchmarks

LOCOMO

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%

AppWorld

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).

---

Integrations — pick your client, copy one line

PowerMem ships first-party plugins and setup guides for the most common AI clients. All of them point at the same backend (HTTP server, MCP server, or local pmem CLI) — no per-client schema rewrites. All agents share the same memory server.

AI agents & IDEs

Claude Code	Cursor	VS Code	Codex	Windsurf	GitHub Copilot
Qoder	OpenCode	OpenClaw	Claude Desktop	Cline

SDKs & apps

App / framework	Details
Python SDK	pip install powermem, see Quick start
LangChain / LangGraph	pip install powermem, see LangChain guide
Go apps	SDKs
Java apps	SDKs
TypeScript apps	SDKs
Any MCP client	powermem-mcp sse (default :8848), see MCP client guide
HTTP REST apps	powermem-server --host 0.0.0.0 --port 8848, see API server

OpenClaw (ClawdBot)

OpenClaw gains long-term memory through the memory-powermem plugin.

openclaw plugins install memory-powermem

Defaults 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).

Full guide: OpenClaw integration.

Claude Code

Fastest path — let Claude Code set itself up

First download the code and enter the directory:

git clone https://github.com/oceanbase/powermem
cd powermem

Then open Claude Code in your terminal and paste this one line:

Read and follow apps/claude-code-plugin/SETUP.md to set up PowerMem memory for Claude Code.

Claude Code reads apps/claude-code-plugin/SETUP.md, asks you for the few required secrets, and wires everything up end-to-end.

Manual setup

Prefer to wire it by hand? See the full walkthrough — environment variables, MCP mode, the remember / recall skills, Windows hooks, troubleshooting, and uninstall — in docs/integrations/claude_code.md.

Cursor, VS Code, Windsurf, GitHub Copilot, Qoder

Recommended setup — let your IDE agent set it up

First download the code and enter the directory:

git clone https://github.com/oceanbase/powermem
cd powermem

Then open the AI agent window in your IDE and paste this one line:

Read and follow apps/vscode-extension/SETUP.md to setup PowerMem

The agent follows apps/vscode-extension/SETUP.md: it prefers a reusable powermem-server HTTP API backend, falls back to MCP-only only when HTTP is unavailable, and configures the current IDE/client instead of unrelated tools.

Manual setup

Prefer to wire it by hand? Use the per-IDE guide:

Client	Details
VS Code	docs/integrations/vs_code.md
Cursor	docs/integrations/cursor.md
Windsurf	docs/integrations/windsurf.md
GitHub Copilot	docs/integrations/github_copilot.md
Qoder	docs/integrations/qoder.md

The same extension also provides Query memories, Add selection to memory, Quick note, and a status-bar Dashboard. See apps/vscode-extension/README.md and the full VS Code guide.

Any MCP client

For Claude Desktop, Codex, Cline, OpenCode, Roo Code, Goose, or any other MCP-compatible client. please use MCP Client mode. First download the code and enter the directory:

git clone https://github.com/oceanbase/powermem
cd powermem

Then open the AI agent window in your MCP client or IDE and paste this one line:

Read and follow apps/mcp-client/SETUP.md to setup PowerMem

The agent follows apps/mcp-client/SETUP.md: it uses powermem-mcp directly, prefers SSE on port 8848, falls back to streamable HTTP or stdio only when needed, and configures only the target MCP client.

Prefer to wire it by hand? Use the Generic MCP client guide. To remove the integration later, follow apps/mcp-client/UNINSTALL.md. Exposed tools: add_memory, search_memories, get_memory_by_id, update_memory, delete_memory, delete_all_memories, list_memories. Full reference: MCP Server. Client-specific notes: Cline, Codex, and OpenCode.

LangChain & LangGraph

pip install powermem langchain langchain-openai

End-to-end runnable demos:

• LangChain healthcare bot
• LangGraph customer service bot

Full framework guide: LangChain and LangGraph integration.

SDKs

Language	Package
Python	pip install powermem (this repo)
Go	ob-labs/powermem-go
Java	ob-labs/powermem-java
TypeScript	ob-labs/powermem-ts

---

Quick start (Python SDK)

Prerequisites: Copy .env.example to .env and set your LLM API key — that is the only required credential. For zero-config local storage, install the seekdb extra (pip install "powermem[seekdb]", or combine it with server / mcp) so the default OceanBase provider can boot embedded seekdb on disk. Without seekdb, set OCEANBASE_HOST to point at a remote OceanBase cluster, or switch to sqlite / postgres. The default embedder is a local all-MiniLM-L6-v2 model (384 dims) that needs no API key and auto-downloads on first use. Need to tune providers or unlock advanced features? Copy .env.example.full instead — it documents every available knob, grouped by component. After install, pmem config init walks you through the same setup interactively. See Getting started.

Install

# Core only (SDK; no optional CLI/server/MCP/seekdb)
pip install powermem

# With CLI (pmem / powermem-cli)
pip install "powermem[cli]"

# With HTTP API server only (powermem-server; does not install seekdb)
pip install "powermem[server]"

# With MCP server only (powermem-mcp; does not install seekdb)
pip install "powermem[mcp]"

# With seekdb for zero-config local storage/embedder
pip install "powermem[seekdb]"

# HTTP API server + seekdb
pip install "powermem[server,seekdb]"

# MCP server + seekdb
pip install "powermem[mcp,seekdb]"

# Common full local install
pip install "powermem[cli,server,mcp,seekdb]"

SDK

Run 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.

CLI (pmem, 1.0+)

pmem memory add "User prefers dark mode" --user-id user123
pmem memory search "preferences" --user-id user123
pmem shell                           # interactive REPL

Full reference: CLI usage.

HTTP API server + Dashboard

Uses the same .env as the SDK. Dashboard is served under /dashboard/.

powermem-server --host 0.0.0.0 --port 8848

Open the dashboard at http://localhost:8848/dashboard/ to browse memories, view analytics, and monitor system health. See the Web Dashboard Guide for a complete walkthrough.

Docker / Compose: see API Server and Docker & deployment. The official image is oceanbase/powermem-server:latest.

---

Capabilities

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

---

Docs

• Getting started — install, .env, and first Memory usage
• Configuration — settings model, storage backends, environment variables
• Web Dashboard — visual interface for memory inspection, analytics, and system monitoring
• Architecture — major components, storage layout, and retrieval flow
• API & services — REST, MCP, HTTP server, and Python-facing APIs
• CLI — pmem commands, interactive shell, backup and migration
• Multi-agent — scopes, isolation, and cross-agent sharing
• Integrations — LangChain and other framework wiring
• Ecosystem integrations — AI clients & IDEs (Claude Code, …)
• Docker & deployment — images, Compose, and running the API server
• Development — local setup, tests, and contributing

More topics: Sub stores, guides index.

Examples

• 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).

Release highlights

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

Support

• GitHub Issues
• GitHub Discussions

License

Apache License 2.0 — see LICENSE.