Memori

Memori introduces an agent-native memory infrastructure that enables building structured, long-term memory directly from operational traces. Key features include: * Capturing execution paths and tool results * Recording workflow steps and outcomes * Storing decision-making logic * Classifying chat turns into facts, preferences, and rules * Selective recall of relevant context This system allows memory to be generated from an agent's actual activities, providing a persistent state for production systems. It automatically captures conversations, classifying information and providing control over data storage duration and location. When context is required for a prompt, Memori efficiently retrieves only the relevant information across discussions and documents, eliminating the need to manage additional services. Memori enhances search accuracy by automatically enriching queries with semantic context when language is ambiguous, preventing excessive token costs. Each retrieved result includes a clear explanation for its inclusion, allowing teams to trace relevance by entity, time, and source. Benchmark results show 81.95% accuracy on LoCoMo while using only 1,294 tokens per query, representing approximately 5% of full-context cost and saving users over 95% on inference expenses. This robust framework is suitable for developers and engineering teams building sophisticated conversational applications or complex operational agents that require efficient and accurate memory management.