LongMemEval

Conversational recall benchmarks such as LoCoMo and LongMemEval measure question-answering ability over conversation history, rather than whether memory has changed the actual behavior of tool-using agents.

mem0 publishes benchmark scores of 92.5 on LoCoMo and 94.4 on LongMemEval (April 2026 figures, measured on its hosted platform, with a note that the OSS direction is similar).

awareliquid achieves Recall@5 of 96.0% on LongMemEval, with zero LLM calls in the retrieval process.

In the LongMemEval test, Hy Memory still scored 45% when using a larger retrieval limit (40 memories, about 7,900 characters), indicating that information loss occurs in the writing stage where conversations are converted into facts rather than in the retrieval stage.

In the LongMemEval benchmark, when using Gemini 3.8 Flash as the writing model, Hy Memory 2.0.1 answered only 43% of 60 questions correctly, while Aura 1.62.0 answered 95% correctly

LongMemEval is a public benchmark for evaluating long-term memory systems, constructing chat histories of about 50 conversation segments, with questions divided into dimensions such as knowledge updates, temporal reasoning, multi-session reasoning, assistant memory and user memory, and judging performed by an LLM via semantic comparison based on the official prompt.

LoCoMo focuses on fact recall and consistency in multi-turn long conversations, while LongMemEval covers a broader range of long-term memory tasks, including temporal reasoning, conflict resolution, and forgetting verification, and both are used in MAQS-Way as measurable verification means.

Graphiti's contradiction-triggers-invalidation mechanism is exemplary (old edges are retained and marked with expired_at), and its evaluation story is the most complete of the three (unit tests + end-to-end tests + LongMemEval oracle), but it has gaps in the audit sink and unscoped search.

MAQS-Way is positioned as the pathway layer between the policy layer and the audit layer, comprising 8 principles (P1–P8), a data model based on OMPS, lifecycle management, and measurable verification means (LoCoMo, LongMemEval, deletion lattice, negative tests).

LongMemEval is a benchmark for evaluating long-horizon memory capability, in which multi-hop tasks require the system to reason across multiple memory fragments to answer a single question.

On the LongMemEval multi-hop task, strict recall@5 performs as follows: 78.8%->85.2% with 5 chunks, 87.2%->92.8% with 10 chunks, and flat at 20 chunks.

On the LongMemEval multi-hop task, increasing the number of retrieved chunks improves recall, but it enters a plateau at 20 chunks, indicating that the bottleneck is not retrieval width but the inherent limitation of vocabulary matching.

LoCoMo, LongMemEval, and MemoryAgentBench provide measurable R@k, temporal reasoning, and abstention metrics; OWASP Agentic ASI06 specifically targets memory and context poisoning.

Memory benchmarks such as LongMemEval and LoCoMo measure whether the answer correctly reaches the model and how many context tokens are consumed, not document retrieval itself.

MemFold's primary training scenario is persona-driven personalization tasks, but it can transfer to PrefEval and LongMemEval without target-domain training.

The developers claim that Hindsight has achieved state-of-the-art performance on the LongMemEval long-term memory evaluation benchmark.

LongMemEval is an academic benchmark specifically designed to evaluate the long-term memory capabilities of large language models, simulating real-world scenarios spanning hundreds to thousands of conversation turns.

APDMem was tested on the LongMemEval benchmark, achieving strong performance on long-context memory reasoning tasks while accessing only 8% of the full conversation.

LongMemEval is a benchmark dataset specifically designed to evaluate LLMs' long-term memory capabilities, covering tasks such as single-hop factual retrieval, multi-hop reasoning, temporal ordering, and contradiction detection.

On LongMemEval, CTWM achieved a 24.48% token reduction while maintaining overall accuracy at the same level.