Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Synthesizing Procedural Memory: Challenges and Architectures in Automated Workflow Generation

Nishant Gaurav, Adit Akarsh, Ankit Ranjan, Manoj Bajaj

TL;DR

The paper addresses how to endow LLMs with the ability to autonomously synthesize production-grade procedural memory in the form of executable code, using a cross-service Outlook–OneDrive case study. It introduces a four-gap framework—Discovery, Verification, Decomposition, and Scaling—driven by hypothesize, probe, and code, and operationalizes it via Dynamic Model Context Protocol, a probe phase, intrinsic derivation, and Linear State Anchoring with external memory. Key contributions include active tool discovery through semantic search, grounding tool schemas with a Probe routine, replacing reliance on large search trees with intrinsic reasoning and a writeable plan counter, and proposing memory/persistence strategies for production reliability. The approach provides a practical blueprint for turning conversational agents into autonomous workflow architects capable of generating robust, scalable automation at scale.

Abstract

While CodeMem establishes executable code as the optimal representation for agentic procedural memory, the mechanism for autonomously synthesizing this memory from a blank slate remains underexplored. This paper operationalizes the transition of Large Language Models from passive tool-users to active workflow architects. Through a high-fidelity case study of a cross-service orchestration task involving Outlook and OneDrive, we identify and address four structural bottlenecks in automated skill generation: the Discovery Gap involving navigation of large tool registries, the Verification Gap regarding grounding tool response structures, the Decomposition Gap which replaces inefficient search with Linear State Anchoring, and the Scaling Gap focused on concurrency and persistence. We demonstrate that by enforcing a scientific methodology of hypothesize, probe, and code, agents can autonomously write robust, production-grade code skills.

Synthesizing Procedural Memory: Challenges and Architectures in Automated Workflow Generation

TL;DR

The paper addresses how to endow LLMs with the ability to autonomously synthesize production-grade procedural memory in the form of executable code, using a cross-service Outlook–OneDrive case study. It introduces a four-gap framework—Discovery, Verification, Decomposition, and Scaling—driven by hypothesize, probe, and code, and operationalizes it via Dynamic Model Context Protocol, a probe phase, intrinsic derivation, and Linear State Anchoring with external memory. Key contributions include active tool discovery through semantic search, grounding tool schemas with a Probe routine, replacing reliance on large search trees with intrinsic reasoning and a writeable plan counter, and proposing memory/persistence strategies for production reliability. The approach provides a practical blueprint for turning conversational agents into autonomous workflow architects capable of generating robust, scalable automation at scale.

Abstract

While CodeMem establishes executable code as the optimal representation for agentic procedural memory, the mechanism for autonomously synthesizing this memory from a blank slate remains underexplored. This paper operationalizes the transition of Large Language Models from passive tool-users to active workflow architects. Through a high-fidelity case study of a cross-service orchestration task involving Outlook and OneDrive, we identify and address four structural bottlenecks in automated skill generation: the Discovery Gap involving navigation of large tool registries, the Verification Gap regarding grounding tool response structures, the Decomposition Gap which replaces inefficient search with Linear State Anchoring, and the Scaling Gap focused on concurrency and persistence. We demonstrate that by enforcing a scientific methodology of hypothesize, probe, and code, agents can autonomously write robust, production-grade code skills.
Paper Structure (21 sections, 1 figure)

This paper contains 21 sections, 1 figure.

Table of Contents

  1. Introduction
  2. The Discovery Gap: Dynamic Tool Identification
  3. The Context-Scale Trade-off
  4. Mechanism: Semantic Intent to Registry Query
  5. Iterative Refinement
  6. The Necessity of Semantic Search
  7. The Verification Gap: The "Probe" Methodology
  8. The Hallucination of Shape
  9. The Sampling Mandate
  10. Case Study: Probing the Outlook API
  11. The Decomposition Gap: Intrinsic Reasoning vs. Tree Search
  12. The Saturation of Search Benefits
  13. The Reflection Trap
  14. The Modern Approach: Intrinsic Derivation
  15. Linear State Anchoring: The todos Tool
  16. ...and 6 more sections

Figures (1)

  • Figure 1: Comparative analysis of HumanEval/83. The Search-based agent (Left) gets lost in a "patch loop," while the Intrinsic Reasoning agent (Right) produces a deterministic $O(1)$ solution.