MIA

同类别比较
名称	排名	指标
Canner/vulcan-sql	13,769	语言: TypeScript 星标数: 792 分叉数: 42 简介: Data API Framework for AI Agents and Data Apps
antvis/GPT-Vis	13,771	语言: TypeScript 星标数: 792 分叉数: 97 简介: 🤖 GPT Vision, Visualization for AI Friendly! AntV's entry for AI.
bosun-ai/swiftide	13,773	语言: Rust 星标数: 792 分叉数: 71 简介: Fast, streaming indexing, query, and agentic LLM applications in Rust
google-research/bleurt	13,776	语言: Python 星标数: 792 分叉数: 94 简介: BLEURT is a metric for Natural Language Generation based on transfer learning.
Microck/opencode-studio	13,784	语言: TypeScript 星标数: 791 分叉数: 61 简介: web GUI for securely managing local OpenCode configuration
syedazharmbnr1/claude-chatgpt-mcp	13,792	语言: JavaScript 星标数: 791 分叉数: 103 简介: A Claude MCP tool to interact with the ChatGPT desktop app on macOS
RobustBench/robustbench	13,796	语言: Python 星标数: 790 分叉数: 106 简介: RobustBench: a standardized adversarial robustness benchmark [NeurIPS 2021 Benchmarks and Datasets Track]
chirpz-ai/pandaprobe	13,797	语言: Python 星标数: 790 分叉数: 122 简介: open source agent engineering platform: traces, evals, and metrics to debug and improve your AI agents. Integrates with LangGraph, CrewAI, Claude Agent SDK, and more.

关键参数与规格
参数	值
语言	Python
类别	AI Engineering
子类别	Give agent knowledge
关键词	Datasets, Embeddings, Memory management, Retrieval
星标数	791
分叉数	85
贡献者	3
创建日期	2026-03-28
更新日期	2026-10-15
1 天星标数	0
7 天星标数	-1
排名	13,782
上次排名	13,678
排名变化	-104
7 天星标排名	17,589
上次 7 天星标排名	17,543
7 天星标排名变化	-46
总星标排名	13,477
上次总星标排名	13,460
总星标排名变化	-17

Memory Intelligence Agent

An Agent Memory Framework Designed For Deep Research Agents

🚀 Latest News

[April 14, 2026]: 💡 Released the user-facing streaming TTRL code, supporting dynamic data loading and customizable rewards.
[April 3, 2026]: 🦞 Guess what? OpenClaw skill MIA-v1 has landed on Clawhub. Download now and see the magic of memory for yourself!
[April 1, 2026]: 🌈 Full stack is here. Whole Training and Evaluation Codebase, Models and Datasets have been published. Check them out!

📌 Overview

Are you tired of deep research agents that "understand" everything but "remember" nothing? Does your agent struggle with diluted attention in sea of long-form text? Are you frustrated by reasoning failures caused by noisy, irrelevant memories? Do you find your agent memory crippled by skyrocketing computational costs and the logistical nightmare of ever-growing context histories? Most importantly: Is your agent stuck memorizing "what" the result is, while completely failing to learn "how" to get there? If so, you are witnessing the fundamental bottleneck of current agent memory: an incompetent Planner retrieving memories from bloated memory and using non-comprehensive in-context prompts to guide an unprepared Executor in conducting deep research.

MIA (Memory In Intelligence Agent) is a memory framewoek designed for deep research agents. It is developed by a joint team from the Shanghai Institute of Intelligence (SII) and East China Normal University (ECNU). MIA is a paradigm-shifting framework designed to transform agents from "passive record-keepers" into "active strategists." MIA replaces the chaotic "memory dump" with a sophisticated Manager-Planner-Executor architecture:

The Manager: Ultimate librarian that optimizes memory storage to eliminate bloat.
The Planner: Tactical brain that doesn't just "recall," but evolves its strategy via Continual Test-Time Learning.
The Executor: Precision instrument that interprets and follows complex research blueprints with zero friction.

🌟 Key Highlights

🧬 From Growth to Evolution: MIA moves beyond "ever-growing contexts" to strategy-oriented wisdom. By establishing a collaborative loop between parametric and non-parametric memories, agents autonomously evolve in complex, open-world scenarios.
🔬 Seamless Synergy via RL: MIA breaks the stagnation of static memory through an Alternative Reinforcement Learning paradigm. This ensures the Manager, Planner, and Executor act as one cohesive mind, achieving fluid multi-agent cooperation that prior works lack.
🧠 Dynamic Intelligence: Unlike static models, MIA features Continual Test-Time Learning, allowing the Planner to adapt and optimize its research strategies on-the-fly during inference.

🏆 Performance

📊 Experimental Analysis

Our comprehensive evaluation across multiple benchmarks demonstrates that MIA significantly improves the performance of Deep Research Agents:

Elevating the State-of-the-Art (a & b): Comparative bar charts on LiveVQA (multimodal) and HotpotQA (text-only, sandbox-based Wiki search) reveal thatMIA significantly boosts the performance of current SOTA Large Language Models (including GPT-5.4, Gemini-3-Flash, claude-sonnet-4.6), proving its efficacy in both text and complex multimodal reasoning tasks.
The "Small-to-Great" Leap (c): Utilizing a Qwen-2.5-VL-7B-based Executor, MIA enables this 7B model to achieve a staggering performance breakthrough across 7 diverse datasets. Remarkably, the MIA-enhanced 7B model outperforms closed-source general models (GPT-5.4, GPT-4o and Gemini-2.5-Pro in non-tool-calling settings), achieving performance close to that of Gemini-3-Flash (in non-tool-calling settings). This underscores MIA’s ability to unlock "super-model" intelligence within efficient, smaller-scale parameters.
Superiority in Agent Memory (d): When benchmarked against contemporary SOTA agent memory frameworks using a unified Qwen-2.5-VL-7B Executor, MIA achieves top-tier results across all 7 datasets. These results establish MIA as a new benchmark in memory-augmented architectures, offering unparalleled efficiency and reasoning depth.

🦞 OpenClaw Skills

We also provide two MIA versions of OpenClaw skills in Original Version and Trust-Worthy Version, which not only integrate MIA memory framework, but also include trust-worthy judgment mechanism. Here are the MIA memory and trust-worthy demos. In addition, we also provide the reference code for training the Planner model based on the online-talking with OpenClaw bot in Training Version and the lightweight FP8 model checkpoints in Lightweight Model.

MIA Memory Demo:

Trust-Worthy Demo:

🛠️ Tools

1. Online Text Search 💻

The core implementation is mainly in web_tools/server. Open web_tools/run.sh and configure the Google Search Serper key:

export SERPER_KEY_ID="xxxxx"

Start the run script:

cd web_tools
bash ./run.sh

Service SERVICE_URL/server, method SERVICE_URL/server/search

2. Offline Text Search 📖

The core implementation is mainly in local_search. Refer to the setup instructions in search-r1. This project uses wiki25 local retrieval. Configure the path and start the run script:

cd local_search
bash ./run.sh

Service http://localhost:8001/, method http://localhost:8001/retrieve

3. Image-to-Image Search 🎨

The image search cache used in this project: image_search_cache

⚙️ Environment

conda create -n verl python==3.10.12

Run the install.sh script in the train directory to install dependencies. Flash-attention needs to be installed separately:

wget -nv https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu12torch2.7cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
pip install --no-cache-dir flash_attn-2.8.3+cu12torch2.7cxx11abiFALSE-cp310-cp310-linux_x86_64.whl

🧬 Data Preparation

Training: 🤗 Train

Testing: 🤗 Test, 🤗 TTRL

✨ Two-Stage RL Training

⚡ Executor Training

Our implementation is based on VeRL. Key modifications:

The core interaction implementation is in /Executor-Train/Train/verl/experimental/tool_agent_loop.py.
prompt is defined in /Executor-Train/Train/local_search/prompt.py.
Custom dataset processing (CustomRLHFDataset) and reward score computation (compute_score) are in Executor-Train/Train/local_search/mmsearch.py.
Tool implementations are in verl.tools.search_tool.SearchTool and verl.tools.web_image_to_image_search_tool.WebImageToImageSearchTool.
The run script is at /Executor-Train/Train/local_search/run_mmsearch_grpo.sh.

1. Deploy the local text search tool.

2. Configure /Executor-Train/Train/local_search/mm_search_tool_config.yaml and /Executor-Train/Train/local_search/mmsearch.yaml:

mm_search_tool_config.yaml
- tools[0].config.retrieval_service_url: local search service URL
- tools[1].config.fvqa_train_cache_path, tools[1].config.test_cache_path: image search cache paths for the test and validation sets
mmsearch.yaml
- hydra.searchpath: trainer config path
- data.custom_cls.path: custom dataset code path
- actor_rollout_ref.rollout.multi_turn.tool_config_path: path to the tool config mm_search_tool_config.yaml

3. Deploy Qwen3-32B on Node 1 as Planner & Judger:

export VLLM_USE_FLASHINFER_SAMPLER=0
CUDA_VISIBLE_DEVICES=0,1,2,3 \
vllm serve /your_path/Qwen/Qwen3-32B \
    --tensor-parallel-size 4 \
    --served-model-name "qwen" \
    --gpu-memory-utilization 0.8 \
    --host 0.0.0.0 \
    --port 8002

LLM service at your_url/8002/v1

4. Deploy the Memory-Planner service (can be on the same node as Planner & Judger):

cd Memory-Serve
cd TRAIN_PLANNER

Configure the run script run.sh: set both MEMORY_URL and PLAN_URL to the LLM service deployed in the previous step. To improve training efficiency, memory content and initial plan are collected in advance. Only the replan service is needed here: your_url/5000/replan_train.

5. Configure the training script /Executor-Train/Train/local_search/run_mmsearch_grpo.sh:

JUDGE_URL: judge service, set to your_url/8002/v1
REPLAN_URL: replan service, set to your_url/5000/replan_train
WANDB_API_KEY: WandB API key (optional)
SAVE_CHECKPOINT_DIR: model save path
DATASET_TRAIN: training dataset path
DATASET_VAL: validation dataset path
REF_MODEL_PATH: pretrained model path

6. Start training on Node 2. Navigate to /Executor-Train/Train/:

bash ./local_search/run_mmsearch_grpo.sh

7. Export the model:

python -m verl.model_merger merge \
    --backend fsdp \
    --local_dir /your_path/actor \
    --target_dir /your_path

Download our trained Executor 🤗 here

⚡ Planner Training

Our implementation is based on VeRL. Key modifications:

The core interaction implementation is in /Planner-Train/mem-plan/verl/experimental/multi_turn_loop.py.
prompt is defined in /Planner-Train/mem-plan/local_search/prompt.py.
Custom dataset processing (CustomRLHFDataset) and reward score computation (compute_score) are in /Planner-Train/mem-plan/local_search/mmsearch.py.
The run script is at /Planner-Train/mem-plan/local_search/run_mmsearch_grpo.sh.

1. Deploy the Judger service on Node 1:

export VLLM_USE_FLASHINFER_SAMPLER=0
CUDA_VISIBLE_DEVICES=0,1,2,3 \
vllm serve /your_path/Qwen/Qwen3-32B \
    --tensor-parallel-size 4 \
    --served-model-name "qwen" \
    --gpu-memory-utilization 0.8 \
    --host 0.0.0.0 \
    --port 8002

2. Deploy the Executor service on Node 2.

Deploy the trained Executor:

export VLLM_USE_FLASHINFER_SAMPLER=0
CUDA_VISIBLE_DEVICES=0,1,2,3 \
vllm serve /your_path/Executor \
    --tensor-parallel-size 4 \
    --served-model-name "qwen" \
    --gpu-memory-utilization 0.8 \
    --host 0.0.0.0 \
    --port 8002

Navigate to /Serve/Train_Planner and configure the run script serve.sh:

AGENT_URL: Executor service URL
SERVICE_URL: offline text search service URL
TEST_CACHE_DIR: image-to-image search cache path
MAX_LLM_CALL_PER_RUN: maximum number of interaction rounds between Executor and tools

Start the service:

bash serve.sh

3. Configure the training script /Planner-Train/mem-plan/local_search/run_mmsearch_grpo.sh:

JUDGE_URL: judge service, set to your_url/8002/v1
PLAN_URL: Executor service for plan responses, set to your_url/5000/plan
REPLAN_URL: Executor service for replan responses, set to your_url/5000/replan
WANDB_API_KEY: WandB API key (optional)
SAVE_CHECKPOINT_DIR: model save path
DATASET_TRAIN: training dataset path
DATASET_VAL: validation dataset path
REF_MODEL_PATH: pretrained model path

To improve training efficiency, memory content and image caption are collected in advance.

4. Start training on Node 3. Navigate to /Planner-Train/mem-plan/:

bash ./local_search/run_mmsearch_grpo.sh

5. Export the model:

python -m verl.model_merger merge \
    --backend fsdp \
    --local_dir /your_path/actor \
    --target_dir /your_path

Download our trained Planner 🤗 here

🔍 Inference

💡 TTRL (Development)

This section introduces the TTRL code used for developer performance validation. Its core implementation is based on an improved veRL framework, which first loads the policy model (i.e., the Planner), then reads the test set, and synchronously completes the following during a single-round (1-epoch) inference process: outputting validation results, storing explicit memory (workflows compressed by Memory Manager), and updating implicit memory (parameters of the policy model).

1. Deploy the Memory Manager & Judger service on Node 1:

export VLLM_USE_FLASHINFER_SAMPLER=0
CUDA_VISIBLE_DEVICES=0,1,2,3 \
vllm serve /your_path/Qwen/Qwen3-32B \
    --tensor-parallel-size 4 \
    --served-model-name "qwen" \
    --gpu-memory-utilization 0.8 \
    --host 0.0.0.0 \
    --port 8002

2. Deploy the Executor service on Node 2:

export VLLM_USE_FLASHINFER_SAMPLER=0
CUDA_VISIBLE_DEVICES=0,1,2,3 \
vllm serve /your_path/Executor \
    --tensor-parallel-size 4 \
    --served-model-name "qwen" \
    --gpu-memory-utilization 0.8 \
    --host 0.0.0.0 \
    --port 8002

Navigate to /Serve/MIA-TTRL and configure the run script serve.sh:

AGENT_URL: Executor service URL
SERVICE_URL: offline/online text search service URL
TEST_CACHE_DIR: image-to-image search cache path
MAX_LLM_CALL_PER_RUN: maximum number of interaction rounds between Executor and tools
MEMORY_URL: Memory Manager service URL
TTRL_SAVE: output path during exploration
PARQUET_PATH: test set path with images (required for image datasets)

In /Serve/MIA-TTRL/call_agent.py, configure the text search method (choose one):

from tool_search_local import *   # offline text search
from tool_serper import *   # online text search

Switch the Python script in /Serve/MIA-TTRL/serve.sh to select the mode (supervised / unsupervised):

python agent_serve_ttrl.py ....   # scenario where Ground-Truth is available after each question
python agent_serve_ttrl_nogt.py ....   # scenario where Ground-Truth is unavailable after each question

3. Configure the script /TTRL/TTRL/local_search/run_mmsearch_grpo.sh (supervised) or /TTRL/TTRL-nogt/local_search/run_mmsearch_grpo.sh (unsupervised):

JUDGE_URL: judge service, set to your_url/8002/v1
MEMORY_URL: memory retrieval service, set to your_url/5000/memory
PLAN_URL: Executor service for plan responses, set to your_url/5000/plan
REPLAN_URL: Executor service for replan responses, set to your_url/5000/replan
MEMORY_BANK_SAVE_URL: service for saving memories to the buffer (current batch exploration not yet complete), set to your_url/5000/memory_bank_save
BATCH_EVALUATE_URL: service for evaluating current batch samples, set to your_url/5000/batch_evaluate
CONSOLIDATE_MEMORIES_URL: service for extracting memories from all buffered samples, set to your_url/5000/consolidate_memories
SAVE_MEMORIES_URL: service for saving all memories, set to your_url/5000/save_memory
WANDB_API_KEY: WandB API key (optional)
SAVE_CHECKPOINT_DIR: model save path
DATASET_TRAIN: dataset path
DATASET_VAL: unused, set to the same as DATASET_TRAIN
REF_MODEL_PATH: initial Planner path

Start the service:

bash serve.sh

4. Start the Planner training on Node 3. Navigate to /TTRL/TTRL/ or /TTRL/TTRL-nogt/:

bash ./local_search/run_mmsearch_grpo.sh

💡 TTRL-streaming (Application)

This section introduces the TTRL-streaming code designed for end users. We decouple the original veRL framework into a server side and a client side: the server loads the policy model and remains on standby; the client handles user interactions and streams input data and user feedback to the server in real time, driving the coordinated update of explicit memory (memory archives) and implicit memory (policy model parameters). Users can customize the feedback mechanism and reward format to flexibly adapt to different application scenarios.

1. Deploy the Memory Manager & Judger service on Node 1:

export VLLM_USE_FLASHINFER_SAMPLER=0
CUDA_VISIBLE_DEVICES=0,1,2,3 \
vllm serve /your_path/Qwen/Qwen3-32B \
    --tensor-parallel-size 4 \
    --served-model-name "qwen" \
    --gpu-memory-utilization 0.8 \
    --host 0.0.0.0 \
    --port 8002

2. Deploy the Executor service on Node 2:

export VLLM_USE_FLASHINFER_SAMPLER=0
CUDA_VISIBLE_DEVICES=0,1,2,3 \
vllm serve /your_path/Executor \
    --tensor-parallel-size 4 \
    --served-model-name "qwen" \
    --gpu-memory-utilization 0.8 \
    --host 0.0.0.0 \
    --port 8002

Open /Serve/MIA-TTRL and configure the run script serve_streaming.sh:

AGENT_URL: URL of the Executor service
SERVICE_URL: URL of the text search (offline/online) service
TEST_CACHE_DIR: Cache path for image-to-image search
MAX_LLM_CALL_PER_RUN: Maximum number of interaction rounds between Executor and tools
MEMORY_URL: URL of the Memory Manager service
TTRL_SAVE: Output path during exploration

For text search settings in /Serve/MIA-TTRL/call_agent.py (choose one):

from tool_search_local import *   # Offline text search
from tool_serper import *   # Online text search

If you encounter out-of-memory (OOM) errors during runtime, you can modify the device for loading embedding in /Serve/MIA-TTRL/agent_serve_ttrl_streaming.py to cpu:

class MemoryProcessor:
    def __init__(self):
        """Initialize the OpenAI client based on model configuration."""
        self.model_name = MODEL_NAME
        bert_path = "/your_path/bert/sup-simcse-bert-base-uncased"
        # self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.device = "cpu" # Disable adaptive selection, always load on CPU

3. Configure the server-side script /TTRL-streaming/run_streaming_ttrl_server.sh:

JUDGE_URL: Judge service, set to your_url/8002/v1
TTRL_AGENT_BASE: TTRL-streaming service, set to your_url/5000
GT_REWARD_URL: Service for receiving reward signals, set to your_url/5000/plan
WANDB_API_KEY: WandB API key (optional)
SAVE_CHECKPOINT_DIR: Model save path
REF_MODEL_PATH: Initial Planner path
data.train_batch_size: batch_size used when updating parameters
actor_rollout_ref.actor.optim.lr: Learning rate when updating parameters
actor_rollout_ref.rollout.n: Number of rollout iterations for the GRPO method

Launch on Node 3:

bash run_streaming_ttrl_server.sh

4. Configure the reward service (user feedback) /TTRL-streaming/:

TTRL-streaming/run_gt_reward_server.sh
- JUDGE_URL: Judge service, set to your_url/8002/v1
TTRL-streaming/gt_reward_core.py: Can be modified to implement custom user feedback/reward patterns

Launch on Node 3:

bash run_gt_reward_server.sh

5. Configure the client-side script /TTRL-streaming/run_streaming_ttrl_client.sh:

JUDGE_URL: Judge service, set to your_url/8002/v1
TTRL_AGENT_BASE: TTRL-streaming service, set to your_url/5000
GT_REWARD_URL: Service for receiving reward signals, set to your_url/5000/plan
REF_MODEL_PATH: Initial Planner path
TTRL_SERVER: URL of the veRL server, set to http://127.0.0.1:8765
DATASET_TRAIN: Dataset path
SAVE_STEP: Save interval for the policy model
CLIENT_BATCH_SIZE: Client parallelism count; must be divisible by data.train_batch_size (the minimum number of samples required for one update) set on the server side

Launch on Node 3:

bash run_streaming_ttrl_client.sh

⚖️ License

Released under the MIT License.

🎓 Contributors

PhD Student: Weicheng Meng, Yu Cheng, Zhihang Lin

Student Leader: Jingyang Qiao

Professor: Zhizhong Zhang, Xin Tan, Jingyu Gong, Zhaoxia Yin

Projector Leader: Yuan Xie

🎯 To-Do List

We also plan to release the following next versions:

High-Efficiency Version
Trust-worthy Version

记录了多少星标和分叉？

仓库排名是多少？

记录的语言和类别是什么？

哪个仓库排名最接近？

🚀 Latest News

📌 Overview

🏆 Performance

🦞 OpenClaw Skills

🛠️ Tools

1. Online Text Search 💻

2. Offline Text Search 📖

3. Image-to-Image Search 🎨

⚙️ Environment

🧬 Data Preparation

✨ Two-Stage RL Training

⚡ Executor Training

⚡ Planner Training

🔍 Inference

💡 TTRL (Development)

💡 TTRL-streaming (Application)

⚖️ License

🎓 Contributors

🎯 To-Do List