Human-Inspired Memory Architecture for LLM Agents

Authors: Doga Kerestecioglu, Alexei Robsky, Clemens Vasters et al.

arXiv · PDF

Summary

The authors build a biologically-inspired memory architecture for LLM agents with six mechanisms: sleep-phase consolidation, interference-based forgetting, engram maturation, reconsolidation on retrieval, entity knowledge graphs, and hybrid retrieval. They introduce a synthetic calibration method to set pipeline thresholds without exposing evaluation data. On a VSCode issue-tracking dataset, deduplication-based consolidation achieves 97% retention precision with 58% storage reduction; on LongMemEval personal chat (streaming evaluation with 475 sessions, ~540K turns), the pipeline matches raw retrieval accuracy at 200K-token budget while enabling an accuracy/storage trade-off curve.

Main takeaways:

• Six cognitive mechanisms address failure modes of naive memory accumulation: consolidation, forgetting, maturation, reconsolidation, knowledge graphs, hybrid retrieval
• Synthetic calibration derives all thresholds without benchmark exposure, avoiding evaluation leakage
• On VSCode issues: 97.2% retention precision, 58% store reduction, +21.8 pp over baseline
• On LongMemEval (streaming, 540K turns): matches raw retrieval accuracy (70.1% vs. 71.2%) at 200K budget with tunable accuracy/size trade-off
• At smaller scale (50 sessions), dedup consolidation yields +13.3 pp improvement in preference recall

Relevance

Loosely relevant. This is about persistent memory management across long interactions, not about how personas or behaviors are installed. However, the consolidation/forgetting/retrieval mechanisms could inform how persona-specific information is stored and retrieved if I ever need to model long-horizon persona consistency or memory effects on marker leakage.

Threat model

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses failure, evaluation, benchmark.

Abstract

arXiv:2605.08538v1 Announce Type: new Abstract: Current LLM agents lack principled mechanisms for managing persistent memory across long interaction horizons. We present a biologically-grounded memory architecture comprising six cognitive mechanisms: (1) sleep-phase consolidation, (2) interference-based forgetting, (3) engram maturation, (4) reconsolidation upon retrieval, (5) entity knowledge graphs, and (6) hybrid multi-cue retrieval. Each mechanism addresses a specific failure mode of naive memory accumulation. We introduce a synthetic calibration methodology that derives all pipeline thresholds without benchmark data exposure, eliminating a common source of evaluation leakage. We evaluate on two benchmarks. First, a VSCode issue-tracking dataset (13K issues, 120K events) where deduplication-based consolidation achieves 97.2% retention precision with 58% store reduction (+21.8 pp over baseline). Second, the LongMemEval personal-chat benchmark where we conduct the first streaming M-tier evaluation (475 sessions, ~540K unique turns). At a 200K-token context budget, our pipeline matches raw retrieval accuracy (70.1% vs. 71.2%, overlapping 95% CI) while exposing a tunable accuracy/store-size operating curve. At S-tier scale (50 sessions), dedup-based consolidation yields a +13.3 pp improvement in preference recall.