Less Context, More Accuracy: A Bi-Temporal Memory Engine for LLM Agents Where a Lean Retrieved Context Beats the Full History

A new open-source memory engine called Engram outperforms the common practice of feeding an LLM agent its entire conversation history, achieving higher accuracy while using roughly eight times fewer tokens, according to research posted on arXiv [1]. The system, detailed in a paper submitted on 5 June 2026, addresses a core limitation of large language model agents: their tendency to forget information across sessions [1]. The standard workaround of replaying the full history into the prompt is described as expensive, slow, and increasingly inaccurate as distractors accumulate [1]. Engram is built on a bi-temporal data model with a dual-process architecture. A fast write path appends lossless episodes without placing an LLM on the critical path, while an asynchronous process extracts atomic facts, builds a bi-temporal knowledge graph, and resolves contradictions by invalidating rather than deleting old facts, preserving a full provenance chain [1]. On the LongMemEval_S benchmark, a set of 500 questions, Engram's lean configuration scored 83.6% accuracy when answering from a retrieved context of approximately 9,600 tokens [1]. The full-context baseline, which processes the entire history of roughly 79,000 tokens, scored 73.2% [1]. The 10.4-point gain was statistically significant, with a McNemar test p-value below 10^-6, and the lean configuration made zero errors across the 500 questions [1]. The paper notes that the accuracy gain requires a hybrid read path that fuses dense, lexical, graph, and recency signals, as using extracted facts alone loses recall while adding retrieved chunks recovers necessary detail [1]. The researchers also contribute a neutral, in-repo evaluation harness with the official category-specific judge baked in, and they publish raw per-question logs [1]. The paper documents measurement-integrity pitfalls that can silently distort memory benchmarks, including prompt truncation, home-grown judges, and full-history leaks [1]. Every reported number ships with a command to reproduce it [1].