VerMem: Unified memory management with local and global verifiers¶

Core mechanism reproduction. Paper conclusions, local checks and unreproduced scope are separate.

Paper information¶

Field	Details
Paper	Verifiable Memory: Learning Unified Memory Management with Local and Global Verifiers for Large Language Model Agents (arXiv 2608.03137)
Institution	Sun Yat-sen University / Xiaolong Sun
First public release	2026-08-04 (arXiv v1)
Author code	Released code: Sun-SYSU-24/VerMem.
Local adapter or method key	`vermem`
Local reproduction code	`src/auto_research/agent_research/`

Original paper summary¶

Background and main changes¶

Long-term memory, active context and episodic history are often optimized separately, and trajectory reward cannot verify one memory action. VerMem uses one policy over all three states and seven atomic operations. A local verifier checks state transitions and a global verifier checks evidence consistency.

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

Core equations¶

\[ R=R_{task}+\lambda_lV_{local}(m_t,a_t,m_{t+1})+\lambda_gV_{global}(M_T,\tau)-\lambda_cC(a_t),\quad \max_\pi\mathbb E_\pi[R]. \]

Offline and online results¶

The paper reports best results on most metrics over five benchmarks and two backbones, and the best efficiency–performance frontier at a controlled online-token budget. No production A/B.

Local reproduction¶

Explicitly maintain long-term memory, active context and episodes. Run retrieve/add/restore and count local and global verifier calls separately.

PlanBench mini-suite, 120 episodes, seed 42: joint success 1.0000, average cost 0.6100, with nonzero telemetry for paper-specific operations.

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

See stable batch metrics.

Scope¶

The deterministic public mini-suite verifies core state updates at matched budgets, not paper-scale models, multi-GPU RL, private environments or full benchmarks. Local changes are not paper gains.