graph-memory

Why we included this project

Agentic coding tools like OpenClaw and DeepSeek Harness share a common problem: each new session starts with no memory of the last, and long conversations bloat the context window. Graph Memory tackles this by storing conversation knowledge as a typed graph, where tasks, skills, and events become nodes linked by edges like USED_SKILL and SOLVED_BY. When a new question comes in, it pulls only the relevant subgraph rather than replaying the whole transcript, which the project says cuts context use by about 75%. Recent user turns stay verbatim, and the memory survives restarts, so a later session can automatically recall what an earlier one learned. For teams building agents that need durable, explainable memory instead of an ever-growing prompt, this plugin is worth a close look.