GraphMem: Self-Evolving Graph-Based Memory for Production AI Agents

Memory systems constitute a critical bottleneck in deploying production-grade AI agents. Existing solutions fail to capture structural relationships, handle contradictory facts, or provide theoretical guarantees. We present GraphMem, a self-evolving graph-based memory architecture with five key innovations: (1) LLM-powered knowledge extraction with exhaustive entity, relationship, alias, and temporal validity capture; (2) multi-signal entity resolution combining lexical, embedding, and LLM-based matching; (3) self-evolving memory with PageRank-based importance scoring and LLM-driven decay; (4) priority-based conflict resolution, the only system with explicit newer-over-older fact handling; and (5) context engineering with relevance-weighted assembly. On the MemoryAgentBench benchmark, GraphMem achieves state-of-the-art results, including eighty percent accurate retrieval, representing a fourteen point nine percentage point improvement over HippoRAG version two, and forty three point three percent conflict resolution, representing a thirteen point eight percentage point improvement. We provide formal guarantees including bounded memory growth under decay, where the number of active edges in memory grows proportionally to the decay rate multiplied by the effective time horizon; query complexity that grows proportionally to the number of nodes multiplied by the embedding dimension, plus the maximum neighborhood expansion raised to the power of the search depth; provably correct conflict resolution that deterministically prefers newer facts over older ones; and a context assembly procedure that achieves a near-optimal relevance-weighted context selection comparable to the best possible greedy strategy. GraphMem is released as production-ready open-source software supporting Neo4j, Turso or SQLite, and Redis backends.