DSPyWeekly Issue No #22

ACM launches CAIS 2026, a new conference on AI and agentic systems | EurekAlert!

“The field at large is still guessing how we can engineer AI software systems that are reliable,” said Omar Khattab, Steering Committee member, Assistant Professor at MIT, and creator of DSPy. “If you want to put a model in a loop, surrounded by control flow and other models, that’s a whole lot harder than talking to ChatGPT yourself. We need to make this an engineering discipline, not a collection of hacks.”

How to Build Your Own Custom LLM Memory Layer from Scratch?.

Step-by-step guide to building autonomous memory retrieval systems

DSPy + Local Nemotron-30B: Quantized LLM Training for Pharma / MedTech - YouTube

In pharma/medtech, AE vs. PC classification determines routing to Safety vs. Quality systems. Misclassification creates queue errors and rework. Most manual effort sits in categorization within the correct flow—assessors constantly reference SOPs, process maps, and prior cases. Converting this to a DSPy-trained model with a simple HITL check can reduce upfront review effort by ~50–90%. No benchmarks, no model chasing—just a practical DSPy training loop on realistic (synthetic) data. Problems like this don’t require “mega-super-ultra” models.

dspy.RLM codebase analyzer

dspy.RLM analyzing a code base with a rules file

Evolving Node Prompts with GEPA on Open-Source LLMs

Now we face the challenge of improving prompting quality for our open models. We basically have two options: iterate manually or run some automation. This is where GEPA comes to our rescue.

The Potential of RLMs

The LLM will use the REPL to filter, chunk, and sample the long context as needed to complete its task. It will use the sub-LLM function to task new LLM instances to explore, analyze, or validate the context. Eventually, the sum of the LLM’s findings will be synthesized into a final answer. That’s it. That’s an RLM.

Built a multi-agent system on top of Blender's Python API and sharing architecture learnings for the MCP discussion - Contributing to Blender - Uses DSpy

I’ve been working on a project called 3D-Agent — a multi-agent system that operates Blender through bpy. The agent reads the scene, plans what to do, writes and executes Python code, then verifies the result by taking a viewport screenshot before moving to the next step.

dspy.RLM - Recursive Language Models in DSPy - Issac Miller

In this conversation, Ankur Gupta interviews Isaac Miller, a core contributor to DSPy, focusing on Recursive Language Models ... | Channel: Information Shelf

myanvoos/coview

Small CLI tool to co-view a research paper or long document with your agent using DSPy's RLM. | Language: Python | License: MIT License

alishivani666/RLMOptimizer

Autonomous prompt optimizer for DSPy. An RLM that runs your program, analyzes what went wrong, and iterates on the prompts until they work. | Language: Python | License: MIT License