LLM agents can now learn continuously from experience, without a single parameter update.
Our latest research at Megagon Labs introduces a memory-driven framework that enables agentic systems to improve from feedback without the cost, inflexibility, and opacity of fine-tuning.
Why Fine-Tuning Falls Short
Traditional approaches to improving LLM performance rely on parameter updates, which are expensive, requiring retraining for every new task, and offering limited interpretability. We propose a different path: memory-augmented learning that leverages structured self-critique grounded in labeled data.
A Framework Built on Episodic and Semantic Memory
Our system uses two types of memory inspired by human cognition:
- Episodic memory stores instance-specific critiques, capturing what went wrong (or right) in particular cases
- Semantic memory distills patterns into reusable, task-level guidance
When the agent encounters new tasks, it retrieves relevant past experiences and learned principles, enabling adaptation without touching model parameters.
Key Results
Testing across diverse tasks (medical QA, preference prediction, reasoning challenges) and six models (GPT-4o-mini, GPT-5, GPT-oss-20b, Llama-4-Scout, Qwen3-235B-instruct, and Qwen3-235B-thinking), our best-performing strategy achieved:
- +8.1 percentage points over zero-shot baselines
- +4.6 percentage points over RAG-only approaches
- 31.95% reduction in inference-time thinking tokens for reasoning models
The efficiency gains are remarkable: pre-computed critiques substantially reduce computational burden, with models “offloading” reasoning they would otherwise perform independently.
Why Some Models Benefit More
We introduce suggestibility, a novel metric measuring how receptive a model is to external reasoning. Our analysis reveals suggestibility correlates with learning effectiveness over zero-shot (Spearman coefficient=0.402, p=0.008), explaining why certain models benefit dramatically while others show modest gains.
High suggestibility enables learning but raises safety concerns: models that readily accept external reasoning may also be vulnerable to adversarial inputs, a critical tradeoff for future research.
What This Means for Agentic AI
This memory-driven approach offers a lightweight, interpretable alternative to fine-tuning for continual agent adaptation. Rather than treating model improvement as purely statistical parameter optimization, we enable agents to learn reflectively from experience, maintaining performance across evolving tasks without retraining.
For teams building AI agents, this research demonstrates that effective adaptation doesn’t require expensive model updates. Smart memory architecture can deliver both accuracy gains and efficiency improvements.
Read the paper here.
Code & datasets: github.com/megagonlabs/critique-learning
Research by Jackson Hassell, Dan Zhang, Hannah Kim, Tom Mitchell, and Estevam Hruschka
Presented at CAIS ’26 (ACM Conference on AI and Agentic Systems)
