MemoraX's Memory-Model-Driven RSI: 20% Higher Coding Pass Rate, 98% Lower Cost

机器之心 · wechat · 2026-10-10

Machine Heart publishes a technical piece from MemoraX AI on how memory underpins recursive self-improvement (RSI) in agents. Founded by Hao Jianye, former head of Huawei's decision and reasoning lab, the company focuses on learnable, evolvable agent memory models.

Context

OpenAI says it has hit its "automated AI research intern" milestone and aims for a true automated AI researcher by March 2028; Anthropic says Claude led roughly 26% of its internal AI R&D work as of August. RSI is now on top labs' agendas.

Method: Memory Model-Driven RSI

A dedicated memory model distills and updates procedural memory from agent trajectories and online feedback, which the agent then applies to real tasks, forming a loop of extraction, application, verification and memory update.
Memory writing is modeled as a learnable policy optimizing future task payoff, trained with agentic RL; memory content is decoupled from access structure for dynamic retrieval; the work also tackles memory credit assignment.

Results

Long-horizon coding: 13 industrial tasks across 5 industries, 78 runs, 335+ hours total, pass rate up 20%+ and main coding agent token use down 20%+; in the hard AccessSentinel case, average pass rate rose 45%+ and token use fell 66%+.
Financial analysis: on the FinRank benchmark, 30+ co-iterations brought evidence-chain recall near 100%, accuracy up nearly 20% vs Codex and cost down nearly 98%; after 36 RSI rounds, recall hit 96.49% and accuracy rose from 33.33% to 73.33%.
Math reasoning: on the IMO-level Matilda 16x16 combinatorial problem, the agent found a verified 21-rectangle optimum in about 4 hours and 600+ model calls, 13% faster and 98% cheaper than OpenAI Codex.

Original post →

MemoraX's Memory-Model-Driven RSI: 20% Higher Coding Pass Rate, 98% Lower Cost

More from coding & agent