PUBLICATIONS
Selected publications
Research on memory, learning, and intelligent systems from the MemoraX research community.
Squeeze the Soaked Sponge: Efficient Off-policy Reinforcement Finetuning for Large Language Model
ReMix enables reinforcement fine-tuning methods to reuse off-policy data through mix-policy updates, KL-convex constraints, and policy reincarnation. Experiments report competitive reasoning accuracy with substantially fewer rollout samples and training steps.
SWE Context Bench: A Benchmark for Context Learning in Coding
SWE-ContextBench tests whether coding agents can reuse prior experience across related software tasks. It groups issues and pull requests with shared context across repositories and languages, enabling evaluation of context retrieval, task resolution, runtime, and token cost.
MemDev-Bench: Does Memory Help Code Agents? Benchmarking Memory Systems over Real Repository Histories
MemDev-Bench evaluates memory systems for coding agents over real repository histories, covering 555 development tasks and 4,869 historical tasks. Memory improves code localization and editing, but reusing past experience without checking its applicability can also cause failures.
GraphMemix: Query-Aware Evidence Forests for Long-Term Multimodal Agent Memory
GraphMemix organizes multimodal agent memory as query-aware evidence forests, reducing redundant lifecycle costs while recovering complementary evidence. Experiments across four long-term multimodal memory benchmarks establish a new accuracy-cost frontier.
EngramBench: A Capability-Grounded Benchmark for Skill-Evolution Harnesses
EngramBench separates genuine capability abstraction from solution leakage across interactive learning and transfer tasks. Its evaluation shows that procedural memory can guide execution, cut redundant context, and reduce coding time by more than 55%.
From Memory to Guide: Spatio-Temporal Composer for Procedural Coding Memory
The Spatio-Temporal Composer turns retrieved coding experiences into stage-specific, environment-aware Runtime Guides. It improves pass rates on complex EngramBench tasks by 7.2% while reducing token consumption by 32.2%, without parameter updates.
Beyond the Final Embedding: Retrieving Memories from Token-Level Evidence
This work improves memory retrieval through sparse token-level matching, allowing local evidence and whole-memory similarity to work together. The method improves retrieval across eight benchmarks and translates those gains into better long-term memory question answering.
NEAR: Neighborhood-Aware Residuals for Memory Retrieval
NEAR uses information shared by neighboring memories to construct residual representations that distinguish similar records with different facts. The training-free method consistently improves retrieval across eight long-context benchmarks and nine embedding models.