EvoMem gives LLM code search persistent memory to cut redundant exploration
EvoMem stores successful mutation strategies and reuses them across runs and tasks
Evolutionary code search driven by large language models typically discards hard-won mutation strategies after each run, forcing future runs to rediscover the same ideas. EvoMem, introduced by Viktor Volkov and 10 colleagues, tackles this by adding a persistent memory layer that records successful mutation events with provenance. The architecture works in two phases: after each evolutionary run, it extracts and stores structured, task-aware advice; during subsequent runs, it retrieves a small set of relevant instructions based on the current task and program context to guide mutations.
The researchers evaluated EvoMem across geometric optimization, multi-hop question answering, GPU kernel optimization, and related benchmarks. Results show positive average improvements in target metrics or search speed for most settings, though gains vary across tasks. While EvoMem doesn't universally outperform, it demonstrates that persistent memory can reduce redundant exploration and improve strategy reuse in LLM-driven evolutionary search. The paper is available on arXiv (2608.10795) and highlights a promising direction for making code optimization more sample-efficient and transferable.
- EvoMem introduces a persistent memory architecture that captures successful mutation events and reuses them as task-aware advice in future evolutionary runs
- Tested across geometric optimization, multi-hop QA, and GPU kernel optimization benchmarks, showing positive average improvements in most settings
- Reduces redundant exploration by retrieving relevant past strategies based on current task and program context
- Paper available on arXiv (2608.10795) from authors at Skoltech and partner institutions
Why It Matters
Lets LLM-driven code optimization learn across runs, cutting wasted search and accelerating performance gains in real workloads.