Gemini 3.6 Flash: 17% fewer tokens

Data Point

ProGraph matches FullContext on multi-hop memory

Two-layer memory architecture using profile expansion and compression residuals achieves 80.1% on MemHop benchmark without explicit knowledge-graph construction.

Current LLM agent memory systems fail at multi-hop reasoning across sessions. ProGraph trades graph indexing overhead for substring traversal of LLM-generated profiles, reducing API calls while handling 5-hop queries—directly improving agent reliability in long-running applications.

Replaces explicit knowledge-graph memory systems (HippoRAG, A-Mem, Mem0). Requires implementing profile expansion logic and compression residual extraction alongside standard context storage. Worth prototyping now if building agents with multi-session memory; code and benchmark released.

• “profile expansion -- substring-matched traversal of entity names that naturally appear in LLM-written profile narratives, a minimal alternative to explicit knowledge-graph construction”
• “compression residuals -- exact dates, quantities, and named items co-extracted with each profile update at zero extra API cost”
• “ProGraph averages 80.1% on MemHop (matching the FullContext reference) and 78.4% on LoCoMo (exceeding FullContext by 11.3pp)”
• “Long-term memory is essential for LLM agents that interact across sessions, yet current memory benchmarks primarily evaluate single-hop recall, leaving multi-hop association largely unmeasured”