EvoArena + EvoMem: Tracking Memory Evolution for Robust LLM Agents in Dynamic Environments

TL;DR — Almost every agent benchmark assumes the environment holds still. Real deployments do not: tools get updated, software changes, user preferences shift. EvoArena models environment change explicitly, as sequences of progressive updates across terminal, software, and social domains. Current agents are weak here, averaging just 39.6% accuracy. EvoMem, a patch-based memory paradigm that records memory as structured update histories (rather than a flat append-only log), lets the agent reason about how the environment changed. It adds 1.5% on EvoArena, 6.1% on GAIA, 4.8% on LoCoMo, and 3.7% on chain-level accuracy where success requires completing a consecutive sequence of evolving subtasks.

flowchart LR
  E0[Env state t0] --> U1[Update patch 1]
  U1 --> U2[Update patch 2]
  U2 --> En[Env state tn]
  U1 -.records.-> Mem[(EvoMem<br/>patch history)]
  U2 -.records.-> Mem
  Mem --> Agent[Agent reasons<br/>over deltas]
  Agent --> Act[Action on<br/>evolved env]
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  classDef aux fill:#e0e7ff,stroke:#6366f1,color:#312e81
  class E0,En input
  class U1,U2,Agent decision
  class Mem aux
  class Act output

What it is

Two contributions. EvoArena is a benchmark suite that turns environment change into the unit of evaluation: each task is a chain of progressive updates, and the agent must keep its knowledge, skills, and behavior aligned as the world shifts. EvoMem is a memory design that stores update histories as structured patches, so the agent can reason about a state as "the previous state plus these deltas" instead of re-reading a monolithic memory blob. Mechanistic analysis shows EvoMem improves evidence capture: it preserves more of the complete evolving-environment state.

Why it matters

Memory is the load-bearing component for long-horizon agents, and the 39.6% baseline says current agents largely fail when the ground moves under them. The chain-level gain (3.7%) is the result to watch, because chained evolving subtasks are exactly where naive append-only memory degrades — errors in tracking one delta cascade. This is a Tier 2 agentic result with implications for any production agent operating against a live, versioned toolset.

Key points

• EvoArena models change as progressive update sequences across terminal, software, and social domains.
• Current agents average only 39.6% accuracy on evolving tasks.
• EvoMem (patch-based, structured update histories) gains 1.5% on EvoArena, 6.1% GAIA, 4.8% LoCoMo.
• Chain-level accuracy (consecutive evolving subtasks) improves 3.7% — the hardest setting.

Relation to prior wiki

Builds directly on the agent-memory concept page. It shares today's "evolving environment" frame with Evoflux (workflow repair under changing tool catalogs) and EvoBrowseComp (contamination-free benchmark via live-web synthesis). Together these three argue that "assume a static world" is the shared blind spot of the current agent-eval canon. See also agent-benchmarks.

Source: HuggingFace Daily Papers · Paper · raw: raw/huggingface/2026-06-14-evoarena-tracking-memory-evolution-for-robust-llm-agents-in.md