AgentMemBench: A Systematic Benchmark for Evaluating Long-Term Memory Management Strategies in Conversational AI Agents

Researchers introduced AgentMemBench, a benchmark evaluating five long-term memory strategies for conversational AI agents across three datasets and 491 annotated question turns, finding that external key-value store (EKV) memory dominates on all quality metrics, achieving a macro Recall@5 of 0.792, MRR of 0.677, Answer F1 of 0.156, and Faithfulness of 0.354, while in-context windowing, web-augmented memory, graph-based episodic memory, and compression-based summarisation nearly fail on long-range recall tasks. The study, which uses Qwen2.5-7B-Instruct for generation and judging, also reveals that EKV's recall advantage comes with a memory footprint cost of approximately 5,100 tokens versus about 300 for other methods, and it validates two existing memory systems, MemGPT/Letta and HippoRAG, with all code and results released for reproducibility.

arXiv:2608.00009v1 Announce Type: new Abstract: Long-term memory remains a critical bottleneck for conversational AI agents, whose finite context windows cannot support coherent recall across thousands of turns. We present AgentMemBench, a unified, reproducible benchmark evaluating five memory management strategies under identical conditions: in-context windowing ICW , external key-value store EKV , graph-based episodic memory GEM , compression-based summarisation CBS , and web-augmented memory WAM . All are assessed across three public datasets covering long-term multi-session dialogue LoCoMo , task-oriented document grounding MultiDoc2Dial , and persona-grounded multi-session chat MSC , using Recall@k, MRR, nDCG@k, Answer F1, an LLM-judge Faithfulness score, Memory Footprint, and Latency over 491 annotated question turns. Generation and judging both use Qwen2.5-7B-Instruct 4-bit , with greedy decoding for determinism. Our results show that 1 EKV dominates on every quality axis macro Recall@5 0.792, MRR 0.677, F1 0.156, Faithfulness 0.354 ; 2 long-range recall is decisive: on LoCoMo, where the gold turn lies many sessions back, ICW, WAM, GEM, and CBS retrieve almost nothing Recall@5 <= 0.005 while EKV alone reaches 0.573, showing that recency windows, summaries, and entity graphs collapse at long horizons and only dense retrieval scales; 3 CBS is the runner-up on retrieval 0.556 ; 4 WAM equals ICW on in-corpus recall by construction, since external results carry no in-corpus provenance; and 5 EKV's recall advantage carries a footprint cost ~5,100 vs ~300 tokens for ICW/WAM , an explicit accuracy-efficiency trade-off. We additionally evaluate two published memory systems MemGPT/Letta, HippoRAG against the same harness, and release all code, environment, and result artefacts for full reproducibility.