DualMemory

Bin Wen^1,2 · Ruoxuan Zhang³ · Yang Chen^1,2 · Hongxia Xie³ · Lan-Zhe Guo^1,2,*

¹State Key Laboratory for Novel Software Technology, Nanjing University
²School of Intelligence Science and Technology, Nanjing University
³Jilin University
^*Corresponding author

This repository contains the public implementation of the method in:

Aligning Progress and Feasibility: A Neuro-Symbolic Dual Memory Framework for Long-Horizon LLM Agents

Long-horizon agents commonly fail in two different ways:

• Progress Drift: the agent loses track of the global task stage and enters irrelevant or repetitive interaction loops.
• Feasibility Violation: the agent proposes an action that is locally impossible under the current environment state.

DualAlign treats these as separate alignment problems and assigns each one a different memory mechanism:

During inference, Progress Memory guides the candidate action toward the next task stage. Feasibility Memory checks that candidate against the current symbolic state and requests refinement when a hard precondition is violated.

The reported evaluation uses ALFWorld, WebShop, and TextCraft.

Table 1: Main results. Values are reported as mean +/- standard deviation.

Clone this repository and install the method dependencies:

cd public_dualalign
pip install -r requirements.txt

The benchmark environments are not vendored here. Follow the StateAct repository setup for ALFWorld, WebShop, and TextCraft. This includes the ALFWorld package and assets, the WebShop server and data, and the TextCraft package and data. Use the StateAct environment variables for those installations, including ALFWORLD_DATA for the ALFWorld data root. Start the WebShop server before running its S1 or S3 command.

Each environment has an empty configuration template:

dualign/alfworld/alfworld_runs/api_config.yaml
dualign/webshop/our_design/api_config.yaml
dualign/textcraft/our_design/api_config.yaml

Fill in the template locally, or configure the client through environment variables:

api:
  gpt:
    api_key: "YOUR_API_KEY"
    api_base: "https://api.openai.com/v1"
  other:
    api_key: "YOUR_API_KEY"
    api_base: "https://your-provider.example/v1"

export OPENAI_API_KEY="YOUR_API_KEY"
export OPENAI_BASE_URL="https://api.openai.com/v1"  # optional

Keep local credential files and environment variables out of version control.

The complete public workflow is S1 -> S2 -> S3. Run the stages in this order; S3 cannot start until the S1 rules and S2 library exist. The stages are separate:

1. S1 collects trajectories on the disjoint construction task pool and builds Feasibility Memory by mining, verifying, and selecting executable rules.
2. S2 reads the successful S1 trajectories and extracts/indexes the Progress Memory library.
3. S3 loads the S1 rules and S2 library, then evaluates DualAlign on the fixed evaluation task pool.

The published task files determine the task instances and order used by the workflow:

• ALFWorld uses ExpeL-compatible suffix files, tasks/alfworld_s1_tasks_suffix.json and tasks/alfworld_s3_tasks_suffix.json. Each record contains the relative ALFWorld gamefile path.
• WebShop uses tasks/webshop_s1_fixed.json and tasks/webshop_s3_fixed.json, which contain session_idx and the complete target key (asin, query, name, and goal_options). Task count and order are read from the file; --start_task_id and --num_envs are not needed.
• TextCraft uses tasks/textcraft_s1_fixed.json and tasks/textcraft_s3_fixed.json, which contain the seed and the complete target goal and recipe set. Task count and seeds are read from the file; --num_games and --seed_start are not needed.

Run S1 first. This stage collects the 50 construction tasks for each environment and writes the verified Feasibility Memory rule code together with the trajectory logs.

Build all three environments:

bash scripts/run_s1.sh all

Run one environment only:

bash scripts/run_s1.sh alfworld
bash scripts/run_s1.sh webshop
bash scripts/run_s1.sh textcraft

Outputs are written to runs/s1_build/<environment>/ by default. Set DUALIGN_OUTPUT_DIR to use another location.

After S1 finishes, run S2. S2 consumes only the S1 trajectory outputs and writes progress_memory_library.json. It does not run evaluation.

Build all three libraries:

bash scripts/run_s2.sh all

The command above is the complete S2 command. It reads the three S1 output directories and writes one progress_memory_library.json file per environment; it does not evaluate tasks or create a second S1 run. To build only one library, pass alfworld, webshop, or textcraft instead of all.

Run one environment only:

bash scripts/run_s2.sh alfworld
bash scripts/run_s2.sh webshop
bash scripts/run_s2.sh textcraft

After both S1 and S2 complete, run S3. The default script performs three independent evaluation runs (r1, r2, and r3), following the paper's reported evaluation protocol:

bash scripts/run_s3.sh all

Or evaluate one environment:

bash scripts/run_s3.sh alfworld
bash scripts/run_s3.sh webshop
bash scripts/run_s3.sh textcraft

To change the number of independent evaluation runs, set DUALIGN_EVAL_RUNS. For example, run once or five times with:

DUALIGN_EVAL_RUNS=1 bash scripts/run_s3.sh all
DUALIGN_EVAL_RUNS=5 bash scripts/run_s3.sh all

If S1/S2 used a custom output directory, point S3 to it with DUALIGN_S1_OUTPUT_DIR. S3 outputs go to runs/s3_eval/ by default. The WebShop server must be available at http://127.0.0.1:3000.

The wrappers pass the published S3 manifests and the generated S1/S2 artifacts to every environment entrypoint. They fail before an API call when either memory artifact is missing. The task manifests reproduce the paper task pools: 50 S1 tasks per environment, followed by 134 ALFWorld, 100 WebShop, and 100 TextCraft S3 tasks.

The supplied model configuration templates match the models used for the reported runs. Since S1/S2 regenerate LLM-produced rules and progress-memory entries, the public workflow is protocol-reproducible but is not expected to be bit-for-bit identical to the private run. The fixed task files, model settings, stage order, and evaluation protocol are fully specified here.

@inproceedings{wen2026dualmemory,
  title     = {Aligning Progress and Feasibility: A Neuro-Symbolic Dual Memory Framework for Long-Horizon LLM Agents},
  author    = {Wen, Bin and Zhang, Ruoxuan and Chen, Yang and Xie, Hongxia and Guo, Lan-Zhe},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}