GenMem

Generative Symbolic Memory for Self-Evolving Harness

This repository provides the released MemRetriever, MemEvolver, RQ-KMeans codebook, and QA, ALFWorld, and WebShop server LoRAs. Training and offline evolution are maintained in the GenMem code repository.

Released artifacts

Path	Artifact
memr/	Qwen3-8B MemRetriever, v11 RL step 50
meme/	Qwen3-8B MemEvolver, v39 RL step 80
server_qa_lora/	Qwen2.5-7B-Instruct QA/search server LoRA, checkpoint 159
server_alfworld_lora/	Qwen2.5-7B-Instruct ALFWorld Interactor LoRA, checkpoint 65
server_webshop_lora/	Qwen2.5-7B-Instruct WebShop LoRA, checkpoint 66
codebook/	Frozen 48 × 16 × 8 × 8 codebooks; 1024-dimensional float32 centers
prompts.json	Inference system prompts and memory_lookup tool schema
encode_memory.py	Standalone experience-to-SID encoder

Each memory agent includes 80 added SID tokens. Use its supplied tokenizer. Load memr or meme as a subfolder: the repository root is a bundle, not a standalone Transformers model.

Encode your memory into a SID

This only needs the small codebooks and Qwen3-Embedding-0.6B, not the 8B agents:

hf download chuchuxwx/GenMem --local-dir genmem-assets \
  --include 'codebook/*' 'encode_memory.py' 'requirements.txt'
cd genmem-assets
pip install -r requirements.txt
python encode_memory.py --memory 'Inspect a failed tool call, correct its arguments, and verify the result.'

CPU is the default; add --device cuda:0 for a GPU. The encoder downloads the embedding revision pinned in codebook/config.json; use --embedding-model for a matching local snapshot. Output contains sid_codes and four SID tokens in the format <SID_L1_X><SID_L2_Y><SID_L3_Z><SID_L4_W>.

Keep the released embedding model, instruction, pooling, normalization, and codebooks together. The encoder implements the weighted residual quantizer directly; see encode_memory.py and codebook/config.json for the exact settings. Multiple experiences can share a SID. Inputs above 8192 tokens are truncated; precision or rewritten input text can change assignments. The checks in validation.json establish quantizer parity on synthetic vectors, rather than exact reproduction of every historical text-to-SID assignment.

Load MemRetriever or MemEvolver

Using the dependencies installed above:

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "chuchuxwx/GenMem"
role = "memr"  # use "meme" for MemEvolver
tokenizer = AutoTokenizer.from_pretrained(
    repo, subfolder=role, extra_special_tokens={}
)
model = AutoModelForCausalLM.from_pretrained(
    repo, subfolder=role, torch_dtype="auto", device_map="auto"
)

Use the role-specific prompt and tool schema in prompts.json. The runtime must execute memory_lookup against your own bank and return its observation before generating the final answer. MemE reasons again after lookup; the released v39 prompt targets failed trajectories. Preserve SID special tokens when decoding: do not use skip_special_tokens=True.

For the complete retrieval, tool-call, and update loop, follow the offline evolution guide.

Training and evolution code

Task-solving server models

The QA server adapter is a rank-16 LoRA on Qwen2.5-7B-Instruct, trained with Harness-RL for StackPlanner2's Central role. Use the base model's tokenizer and load the adapter on that base, not on the Qwen3-8B memory agents. Follow its download and loading instructions.

The ALFWorld adapter instead runs as Interactor, with Central frozen. The WebShop adapter uses Central in historical evaluation; its guide documents the training-harness provenance limitation. All three use the same Qwen2.5-7B-Instruct base and original tokenizer. Load each task adapter separately and preserve its harness routing.

Full workflow

Use the GitHub repository for the full workflow:

• Embedding and RQ-KMeans
• PT, summary preparation, and CoT-MT
• MemR/MemE GRPO
• Training environments

To train the task-solving server model, use Harness-RL. Keep your memory bank and task data in your own workspace. The paper citation is available in the code repository.

Usage notes and license

Validate generated memories before writing them and retain recoverable snapshots: outputs can contain errors, repetition, or task-specific answers. A SID does not guarantee useful retrieval or exact reconstruction of its source text. Avoid putting secrets or personal data into memory content.

Code, codebooks, and fine-tuned weights use Apache-2.0; see LICENSE and NOTICE. The agents derive from Qwen3-8B, the server adapters derive from Qwen2.5-7B-Instruct, and the encoder uses Qwen3-Embedding-0.6B.