ProGraph: A Two-Layer Memory for LLM Agents Beats Mem0, RAG and HippoRAG

Researchers have introduced ProGraph (Profile-Graph Memory), a two-layer long-term memory architecture for LLM agents. The preprint appeared on arXiv in July 2026. ProGraph scored 80.1% on the new MemHop benchmark and 78.4% on LoCoMo, beating Mem0, A-Mem, HippoRAG and RAG.

Why a new benchmark is needed

MemHop is a benchmark of 1000 questions with reasoning depth from 1 to 5 hops, spread across 10 social-network scenarios, with evidence annotated at every hop. According to the arXiv preprint, existing memory tests mostly check single-hop recall and barely measure multi-hop association, that is, linking facts through a chain of intermediate entities. MemHop fills that gap and makes it possible to evaluate an agent's reasoning depth separately.

How ProGraph works

ProGraph runs on two layers and does not build an explicit knowledge graph. The first layer is profile expansion: a traversal over substrings of entity names, which occur naturally in the text profiles the model writes itself. It serves as a minimal replacement for hand-building a knowledge graph. The second layer is compression residuals: exact dates, quantities and named objects, extracted together with each profile update at no extra API cost (zero extra API cost).

A full-grid ablation (testing every configuration) showed that the two mechanisms are specialized. Removing profile expansion costs 22.6 percentage points on MemHop; not extracting compression residuals jointly costs 8.6 points on LoCoMo. Cross-effects within the same architecture stay below 3 points.

• MemHop benchmark: 1000 questions, depth of 1–5 hops, 10 social-network scenarios
• ProGraph: 80.1% on MemHop (on par with the FullContext reference) and 78.4% on LoCoMo (+11.3 pp over FullContext)
• Profile expansion drives multi-hop performance: −22.6 pp on MemHop when removed
• Compression residuals drive precision: −8.6 pp on LoCoMo without joint extraction
• Beats Mem0, A-Mem, HippoRAG and RAG on both benchmarks

Why the result matters

On LoCoMo, ProGraph beat the FullContext reference by 11.3 points, and on MemHop it matched it. It does this without keeping the entire history in the context window and without building an explicit graph. That means an agent can answer chained questions from compressed profiles and residual facts rather than from the full conversation history.

"Profile expansion enables multi-hop reasoning, while compression

residuals provide extraction precision," the arXiv preprint abstract states.

The authors have open-sourced the MemHop benchmark, the ProGraph implementation and the competitors' baseline implementations, so the reported numbers can be reproduced independently.

What this means

Memory for LLM agents is shifting away from stuffing in the whole context and from heavy knowledge graphs, toward lightweight text profiles with targeted facts. ProGraph shows that a cheap method can match full context on deep questions, and beat it where linking facts across several steps matters.

Frequently asked questions

What is MemHop?

MemHop is a benchmark for testing multi-hop memory in LLM agents: 1000 questions with reasoning depth from 1 to 5 hops across 10 social-network scenarios, with evidence annotated at every step.

How does ProGraph differ from Mem0 and RAG?

Instead of an explicit knowledge graph, ProGraph uses two layers: a traversal over names in text profiles and jointly extracted exact facts (dates, numbers, objects). On the MemHop and LoCoMo benchmarks it beat Mem0, A-Mem, HippoRAG and RAG.

Can the results be reproduced?

Yes. The authors have published the MemHop benchmark, the ProGraph implementation and the baseline implementations, so the measurements can be repeated independently.