A new open-source memory framework enables AI agents to retain long-term knowledge without consuming LLM tokens during memory operations, cutting latency while improving scalability for edge and enterprise AI applications.
Researchers have introduced Zero-Mem, an open-source long-term memory framework for AI agents that eliminates token consumption during memory storage and retrieval. Unlike conventional memory architectures that repeatedly invoke large language models (LLMs) to summarize, update or retrieve stored knowledge, Zero-Mem performs all memory operations without calling an LLM. The language model is used only once—during the final question-answering stage—significantly reducing inference costs and latency while preserving long-term conversational context.
Rather than generating intermediate summaries, Zero-Mem stores original interaction traces as its permanent record. The framework organises these records using two complementary data structures: an entity-context graph, which links related entities across conversations, and a temporal hierarchy, which preserves chronological context and session history. During retrieval, the system dynamically combines information from both structures to locate relevant memories before passing only the required context to the LLM.
The key features are:
- Zero-token memory operations with no LLM calls during storage or retrieval.
- Dual memory architecture combining an entity-context graph and temporal hierarchy.
- Original conversation traces retained as the permanent source of truth.
- Deterministic conflict resolution to improve retrieval reliability.
- Open-source framework designed for scalable long-term AI memory management.
A deterministic calibration stage further filters conflicting or inconsistent information, ensuring that responses remain grounded in verified conversation history instead of generated memory summaries. Because memory management relies on deterministic algorithms rather than repeated LLM inference, the approach removes token overhead while maintaining traceability and reducing computational costs.
Benchmark evaluations on long-memory and long-context question-answering datasets showed that Zero-Mem delivers competitive accuracy while reducing memory-operation time by 57.6% compared with the fastest baseline tested. Since no LLM calls are required for storing or retrieving memories, the architecture also eliminates token costs associated with long-term memory management, making it attractive for production AI deployments where inference expenses scale rapidly.
The framework is intended for AI assistants, autonomous agents, enterprise copilots, customer-service systems and robotics platforms that require persistent memory across long interactions. It is also suitable for edge AI deployments and privacy-sensitive environments because conversation traces remain the source of truth rather than repeatedly generated summaries. As AI systems increasingly rely on persistent context for personalised interactions and autonomous decision-making, Zero-Mem offers an efficient alternative to conventional memory architectures while reducing operational costs and simplifying deployment.
