Skip to content

GaosCode/MemWing

Repository files navigation

MemWing mark

MemWing

Long-term collaborative memory for workplace agents.

MemWing gives workplace agents a governable memory layer: capture Source Events, build Current Truth, surface project context, manage memory lifecycle, and push timely reminders back into collaboration flows.

GitHub repository Python 3.13+ OpenClaw runtime Lark collaboration FastAPI Apache 2.0 license

Quickstart | Control Plane | Runtime Profiles | Architecture | CLI | License | 中文


Why MemWing

Most workplace agents are clever but forgetful. They can answer the current prompt, but lose the thread across project decisions, Lark discussions, tool runs, shifting deadlines, and "we already decided this last week" moments.

MemWing is a long-term collaborative memory system for those agents. It stores authoritative Source Events, derives recallable memory layers, tracks what is current versus historical, and gives humans a Control Plane for review, redaction, lifecycle changes, and proactive push.

It is built for teams that need memory to be useful, explainable, and governable.

Quickstart

The happy path is designed to feel like installing a normal local tool.

brew install memwing
memwing quickstart

Prefer an install script?

curl -fsSL https://raw.githubusercontent.com/GaosCode/MemWing/main/packaging/install.sh | sh
memwing quickstart

memwing quickstart uses the Lite profile by default. It creates local state, links the OpenClaw plugin, configures the context engine, and starts the MemWing runtime.

memwing status
memwing doctor

Control Plane

MemWing is not only a memory API. It includes a Control Plane for humans to inspect and govern what agents remember.

memwing control-plane

Use the Control Plane to:

  • review Memory Items before they become trusted long-term context
  • inspect Source Events behind an answer or push candidate
  • approve, archive, redact, or purge memory with a reason
  • monitor Page Memory, pipeline readiness, and maintenance queues
  • send proactive memory cards back into workplace collaboration flows

Runtime Profiles

Profile Best For Storage Derived Backends Setup
Lite personal trials, demos, OpenClaw users who do not want databases SQLite graph and evidence disabled memwing quickstart
Full Local local evaluation with full retrieval stack Postgres Qdrant + Neo4j / Graphiti memwing quickstart --profile full-local
Production managed infrastructure for teams Postgres managed Qdrant + Neo4j / Graphiti memwing setup --profile production

Lite is intentionally close to consumer software: no Docker, no Postgres, no Qdrant, no Neo4j, and no separate model credentials. By default, model calls go through the agent runtime.

OpenClaw And Lark

MemWing is platform-neutral, but today it has a first-class OpenClaw path and workplace collaboration flows designed around Lark.

With OpenClaw, MemWing provides:

  • a ContextEngine for context assembly, after-turn capture, and compaction
  • conversation hooks for Source Event capture
  • tools for memory search, source lookup, project context, and explanation
  • native-memory compatibility shims for agent workflows

With Lark-style workplace flows, MemWing is designed to remember:

  • project decisions and their reasons
  • changing preferences and deadlines
  • recurring team workflows
  • information that should be resurfaced before it is forgotten

Core vocabulary:

  • Source Event: the authoritative raw collaboration event.
  • Current Truth: the effective fact set preferred at recall time.
  • Page Memory: editable mid-term project or thread summary.
  • Memory Item: governable long-term memory with lifecycle state.
  • Evidence Index: derived retrieval layer over Source Events.
  • Graph Backend: entity and relationship layer for validity and history.
  • Control Plane: human governance surface for memory review and operations.

CLI

# One-command local setup
memwing quickstart

# Start runtime from existing config
memwing start

# Inspect local configuration and health
memwing status
memwing doctor

# Open the memory governance UI
memwing control-plane

# Install or inspect the OpenClaw plugin
memwing openclaw install
memwing openclaw status

# Production config skeleton
memwing setup --profile production

Automation and scripts can also call the Control API through the low-level memwing control ... command family.

Development

Backend setup:

uv sync
uv run pytest
uv run ruff check .

Run the backend locally:

# Start API and Memory Pipeline Runtime together
uv run memwing-runtime

# Or run the entrypoints separately
uv run memwing-api
uv run memwing-pipeline

CLI development:

uv run memwing quickstart --skip-openclaw --no-start
uv run memwing status --no-health
uv run memwing doctor --json

Control Plane frontend:

cd frontend
npm install
npm run dev

OpenClaw plugin development:

cd memwing/integrations/openclaw
npm install
npm run build
npm run smoke

Project Status

MemWing is evolving quickly. The current implementation focuses on the v1 memory runtime, OpenClaw integration, Lite setup, Control Plane APIs, benchmark support, and local-first developer workflows.

License

MemWing is released under the Apache License 2.0. See the root LICENSE file for the full license text.

About

long-term collaborative memory for workplace agents

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors