ProGraph: двухслойная память для LLM-агентов обошла Mem0, RAG и HippoRAG

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

Why a new benchmark was needed

MemHop is a benchmark of 1000 questions with connection depths ranging from 1 to 5 hops across 10 social network scenarios, with evidence annotations at each hop.

According to the arXiv preprint, existing memory tests mostly check single-hop recall and barely measure multi-hop associations — linking facts through a chain of intermediate entities.

MemHop closes this gap and allows the depth of an agent's reasoning to be assessed separately.

How ProGraph works

ProGraph operates on two layers and does not build an explicit knowledge graph.

The first layer is profile expansion: traversal over substrings of entity names that naturally occur in text profiles written by the model itself. This is a minimal replacement for manually building a knowledge graph.

The second layer is compression residuals: exact dates, quantities, and named objects that are extracted together with every profile update, with zero extra API cost.

A full-grid ablation showed that the mechanisms are specialized. Removing profile expansion costs minus 22.6 percentage points on MemHop; not extracting compression residuals jointly costs minus 8.6 points on LoCoMo. Cross-effects within a single architecture stay below 3 points.

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

Why the result matters

ProGraph beat the Full Context baseline by 11.3 points on LoCoMo, and matched it on MemHop — all without keeping the whole history in the context window and without building an explicit graph. This means the agent can answer chained questions by relying on compressed profiles and fact residuals, rather than the full conversation history.

«Profile expansion enables multi-hop reasoning, while compression

residuals provide retrieval accuracy», the arXiv preprint abstract states.

The authors have released the MemHop benchmark itself, the ProGraph implementation, and the baseline implementations of its competitors into open access — meaning the reported figures can be independently reproduced.

What this means

Memory for LLM agents is shifting away from the "stuff the whole context in" approach and heavy knowledge graphs toward lightweight text profiles with pinpoint facts. ProGraph shows that a cheap method can catch up with full context on deep questions — and beat it where linking facts across multiple steps matters.

Frequently asked questions

What is MemHop?

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

How does ProGraph differ from Mem0 and RAG?

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

Can the results be reproduced?

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