Researchers introduce EngramaBench to test AI assistants' ability to remember conversations over months — revealing that graph-based memory beats simple full-context recall on complex reasoning tasks

Researchers at an unnamed institution published EngramaBench, a test suite with 100 multi-session conversations and 150 questions that measure how well AI chatbots remember and reason about information accumulated over time. They compared three approaches: GPT-4o with full conversation history, Engrama (a graph-structured memory system that organizes facts as connected nodes rather than plain text), and Mem0 (an open-source vector-retrieval system that finds similar past messages). GPT-4o full-context scored highest overall (0.6186 out of 1.0), but Engrama outperformed it on cross-space reasoning—connecting facts from different conversation topics—by 2.4 percentage points.

Engrama organizes memories as a structured knowledge graph (a web of connected facts with explicit relationships) rather than a flat text search. This matters because humans don't remember conversations as one long transcript; we link related ideas together. When asked a question that requires reasoning across multiple previous conversations, Engrama retrieved and combined relevant facts more effectively than systems that simply search for similar text, making its answers more accurate for questions like 'Did you tell me about Project X and Project Y using the same technology?'

If you use AI assistants like ChatGPT for work over weeks or months, this research signals a coming shift: future versions will likely use structured memory instead of forcing you to re-paste context every session. For companies building AI customer-service agents or research assistants, choosing between these memory architectures—graph-based vs. simple search—will directly affect whether the AI gives accurate answers about your prior interactions or makes contradictory claims.