CoEvo-Mem: Alternating co-evolution of retrieval routing and memory banks¶

Core mechanism reproduction. Paper results, local mechanism checks and missing scope are distinct.

Paper information¶

Field	Details
Paper	CoEvo-Mem: Co-Evolving Retrieval Policy and Memory Bank for LLM Agents (arXiv 2608.01739)
Institution	Not listed in the paper
First public release	2026-08-03 (arXiv v1)
Author code	No released official code found on 2026-08-08.
Local adapter or method key	`coevo-mem`
Local reproduction code	`src/auto_research/agent_research/`

Original paper summary¶

Background and main changes¶

Optimizing routing or memory in isolation ignores their feedback loop. CoEvo-Mem uses a frozen LLM for route-specific rewrites and priors, with a lightweight residual router adapting online. Task outcomes update routing; trajectories update memory values and graph relationships. Alternate which side is frozen to control nonstationarity.

flowchart LR
 A["Student or agent rollout"] --> B["Paper-specific planning or memory operator"]
 B --> C["Auditable objective or state update"]
 C --> D["Public mini-suite evaluation"]

Key figure from the paper¶

Figure 2: Training workflow and key optimizations. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

\[ q'=q_{LLM}+\Delta_\phi(q),\quad \phi\leftarrow\arg\max J(\phi;M\ \mathrm{fixed}),\quad M\leftarrow\operatorname{Update}(M;\tau,\phi\ \mathrm{fixed}). \]

Offline and online results¶

The authors report state of the art across seven varied benchmarks to validate retrieval-memory co-evolution. No production A/B.

Local reproduction¶

Implements route rewriting, residual routing and alternating router/memory-bank updates. Retrieval keys include the task axis and tool signature to prevent mistaken reuse across tasks.

PlanBench mini-suite, 120 episodes, seed 42: joint success 1.0000, average cost 0.5000. All paper-specific operations record nonzero telemetry.

auto-research agent-eval --method coevo-mem --benchmark planbench-mini --episodes 120 --seed 42
auto-research evolve --model agent --dataset planbench-mini --direction "组合 coevo-mem 与已安装论文算子" --generations 2 --population 4

Stable metrics: batch audit.

Scope¶

The deterministic public mini-suite checks core state updates at matched budgets. It does not reproduce large models, multi-GPU RL, private environments or the full benchmark. Local differences are not paper gains.