adoresever/graph-memory

What it solves

Graph Memory provides a traceable, searchable, cross-session memory system for AI agents. It prevents the need to replay entire conversation histories (which consumes tokens and context window) by instead recalling only the most relevant pieces of knowledge from previous sessions.

How it works

The system converts conversation knowledge into a typed knowledge graph consisting of nodes (Tasks, Skills, and Events) and typed edges (e.g., USED_SKILL, SOLVED_BY, REQUIRES) to preserve relationships.

It uses a dual-path recall mechanism:

1. Exact Path: Combines semantic vector retrieval (with FTS5 lexical fallback) and graph traversal to find specific information.
2. Generalized Path: Matches community summaries to retrieve broader context.

Relevant subgraphs are then ranked using Personalized PageRank (PPR) and injected into the agent's prompt within a defined token budget. It integrates natively with the DeepSeek Harness (DSH) and OpenClaw, using SQLite by default for local-first storage.

Who it’s for

Developers building AI agents that require long-term, durable memory across different sessions without bloating the prompt context with irrelevant history.

Highlights

• Cross-Session Persistence: Knowledge from one session is automatically recalled in another, surviving restarts.
• Token Efficiency: Replaces indiscriminate history replay with a relevant knowledge subgraph, significantly reducing token usage in long workflows.
• ** uma Local-First Approach:** Uses SQLite by default, making it lightweight and requiring no external graph database deployment.
• Native Integration: Integrates directly into the DeepSeek Harness plugin lifecycle rather than using a side-channel.
• Verifiable Memory: Includes tools like gm_status and gm_record for deterministic knowledge recording and state reporting.