Graph Memory

Graph Memory is a memory plugin for the DeepSeek Harness, with a maintained adapter for OpenClaw. It takes over the model-visible history of an agent conversation without deleting the host's event log: by default it keeps the newest five completed user turns, strips completed reasoning and tool traces from future requests, and automatically recalls relevant older or cross-session question-and-answer pairs when they matter.

The graph is a navigation layer, not a replacement for evidence. Each completed turn is written as one self-contained summary, from which subject-predicate-object triples are derived, and the summaries, triples, and detected communities only help locate memories: the original question and final answer remain the stored evidence that travels with any graph hit. Writing one turn costs a single auxiliary model call, while community detection and ranking run locally. Recall uses vector top-k search with a full-text fallback, and invalid extractions are quarantined rather than repaired.

Memory is durable in local SQLite with stable provenance and cross-session, cross-project recall. A reported 20-turn run on GLM-5.2 cut the first-request context at turn 20 from about 57,000 tokens to about 11,000, roughly an 80 percent reduction. Embeddings are optional through OpenAI-compatible endpoints, and without them it falls back to full-text search without blocking the conversation.