MemLoop SFT v6
fine-tuned model for multi-agent systems with feedback support
2 min readupdated
This is a specialized fine-tune of the Qwen3-4B-Instruct model, trained to manage three memory agents (memory, retrieval, answer) and perform a callback function. The model is capable of spotting errors in its own responses or in the work of other agents, generating specific feedback, and sending tasks back for re-execution. The checkpoint is created based on sft-longmemeval training and serves as a starting point for further reinforcement learning (GRPO). This new release will be useful for developers building complex context processing and long-term memory pipelines for LLMs.
Key features
- Three roles: The model is trained to play the roles of memory, retrieval, and answer within a unified memory system.
- Callback mechanism: The ability to analyze errors and return tasks for re-execution with a budget of up to 3 iterations.
- Metric improvements: Shows a quality boost on the LongMemEval and LoCoMo benchmarks when using the feedback mechanism.
- Inference readiness: Full out-of-the-box compatibility with popular vLLM and SGLang engines.
What you can do
- Pipeline integration: Deploy the model via vLLM or SGLang to manage long-term agent memory.
- Self-correction testing: Run scenarios with complex query processing where agents check each other's responses.
- RL experiments: Use this checkpoint as a base to run your own reinforcement learning algorithms (GRPO).
Example
The model responds strictly through a tool_call, including special fields for feedback management:
{"name": "answer_agent", "arguments": {"revisit_agent": "retrieval", "feedback": "Missing key facts from step 2"}}
Read more
DetailsBenchmarks (LongMemEval): 75.00% accuracy with callbacks versus 72.25% without them.
Benchmarks (LoCoMo): 80.84% accuracy with callbacks versus 74.88% without them.
Base model: Qwen/Qwen3-4B-Instruct-2507 (4.41B parameters).
Features: Requires a `rope_theta` configuration of 5000000 from `config.json`.










Comments