News

Memory Arena

Major data update: 28 new systems scored from arXiv papers, GitHub repos, and vendor benchmark pages. The catalog now has 234 memory systems with 50 carrying verified benchmark data across LongMemEval, LoCoMo, LongBench, HotpotQA, RULER, NIAH, and more.

Memory Arena

An open dashboard tracking memory systems, vector stores, and long-context architectures across the major published memory benchmarks. Launches with 234 systems and 14 benchmarks.

arXiv

A comprehensive survey covering modular memory architectures proposes standardized benchmarking under a unified framework, evaluating systems across LongMemEval, LoCoMo, and novel multi-source memory tasks.

arXiv

D-Mem proposes quality-gated and full-deliberation memory write strategies. Achieves 78.4 LLM-judge score on LoCoMo with GPT-4o-mini, outperforming Mem0 baseline of 68.7.

UK Tech News / Pulse2

London-based xmemory raises $4M led by Fly Ventures to build typed, schema-backed memory for AI agents. Founded by Alex Petrov (ex-Yandex), the platform structures memory at write time with deduplication and constraints. Currently in closed beta.

EIN Presswire

Backboard.io posts state-of-the-art on both leading memory benchmarks. LongMemEval 93.4% (467/500, independently verified by NewMathData) and LoCoMo 90.0%, both with GPT-4.1. Full replication repos published on GitHub.

arXiv

HiMem introduces a hierarchical memory architecture using topic-aware event-surprise dual-channel segmentation. Outperforms Mem0 (68.7) and A-MEM (51.9) on LoCoMo GPT-Score, with strongest gains in single-hop (89.2) and temporal reasoning (74.8).

Cognee Blog

Independent evaluation comparing Cognee, Mem0, and Zep/Graphiti on knowledge extraction accuracy, relationship mapping, and multi-hop retrieval. Highlights trade-offs between graph-based and vector-based memory approaches.

Zep

Zep sunsets the self-hosted Community Edition while keeping the Graphiti temporal knowledge graph engine fully open source. Cloud-only for production, with Graphiti as the standalone graph memory option.

arXiv

MIRIX stores both text and screenshot memories for agents. Achieves 85.38% overall accuracy on LoCoMo with strong temporal reasoning (88.4%), while maintaining 59.5% on its own ScreenshotVQA benchmark.

arXiv / MemTensor GitHub

MemOS treats memory as a manageable system resource with the MemCube abstraction. Achieves 77.8% on LongMemEval and 75.8 LoCoMo LLM-Judge score, outperforming Memobase, Mem0, Zep, and Supermemory across all categories.

arXiv / ECAI 2025

Mem0's technical report details the hybrid vector + graph memory architecture. Reports 26% relative improvement in LLM-as-Judge over OpenAI memory, 91% lower p95 latency, and 90%+ token cost savings. Published at ECAI 2025.

arXiv

The sequel to HippoRAG achieves 75.5 F1 on HotpotQA with Llama-3.3-70B, a significant jump over HippoRAG 1's 59.2. Introduces deeper passage integration and more effective online LLM use for associative memory tasks.

arXiv

MiniRAG introduces a lightweight graph-RAG approach designed for small language models (1.5B-4B params). With gpt-4o-mini it hits 68.43% on MultiHop-RAG; even Phi-3.5-mini (3.8B) reaches 49.96%, at 25% of typical storage cost.

arXiv

Titans introduces learnable memory modules (MAC, MAG, MAL) that integrate with attention layers. The MAC variant achieves 98.85% on NIAH Passkey retrieval and 97.9% on RULER across context lengths up to 16K, outperforming Mamba, RWKV, and GPT-4.