Engram trims agent memory instead of replaying full history

A new arXiv paper argues that long-term memory is becoming a core layer for LLM agents, but replaying the whole conversation history is expensive and can reduce accuracy as distractors accumulate. The authors introduce Engram, a bi-temporal memory engine designed to retrieve lean context instead.

The system combines lossless episode logging with asynchronous extraction of atomic facts into a knowledge graph, then retrieves only what appears relevant for the current task.

The work matters because agent memory is often sold as a bigger context window. Engram points toward a more selective approach: store broadly, retrieve carefully, and measure whether less context actually improves accuracy.