SodaMem: New Temporal Graph Memory for LLM Agents

Large Language Model (LLM) agents designed for long-term user assistance, spanning weeks of conversation, face a critical challenge: remembering what is *currently* true, not just what was previously stated. Existing memory solutions like flat RAG diaries often struggle with currency, provenance, and ordered temporal reasoning, making it difficult for agents to maintain an accurate "store of record." A new system called SodaMem proposes an evidence-grounded temporal graph memory to overcome these limitations. It operates by extracting "FactEvents" that include mandatory provenance spans, recording when information was mentioned, when it occurred, and its validity. This is managed through a graph structure with SUPERSEDES, CONTRADICTS, and UPDATES edges, supported by hybrid lexical-dense indexing. When an LLM agent needs to answer a question, SodaMem employs a planner-reader loop. This loop first gathers citable evidence from its temporal graph memory before composing a final, accurate response. Benchmarking on LongMemEval-S shows SodaMem achieving high accuracy at a competitive cost, demonstrating its effectiveness in providing LLM agents with a robust, time-aware memory.