Extract recurring multi-step procedures from episodic traces into a procedural DAG.
npx skills add https://github.com/thistleknot/skills --skill procedural-memory
Please help me install this Agent Skill. Skill: procedural-memory Source: https://github.com/thistleknot/skills/tree/main/procedural-memory Command: npx skills add https://github.com/thistleknot/skills --skill procedural-memory
Procedural-memory automatically turns raw episodic traces into reusable, verifiable procedures so an agent can recall and execute multi-step workflows without manual curation.
Use procedural-memory to extract and index recurring workflows from your session execution traces, then query the resulting procedural DAG for the steps required to reach a goal.
High-intent search queries and answers about installing and using this skill.
To extract reusable procedures from execution traces, you need to mine frequent ordered action patterns using sequential pattern mining and verify them with an LLM scoring rubric before committing them to a procedural DAG. This turns raw episodic data into structured workflows.
Sequential pattern mining for procedural memory identifies recurring ordered action sequences within episodic traces. It uses algorithms like PrefixSpan to discover frequent patterns, which are then scored by an LLM to filter for high-quality procedural candidates before DAG construction.
You query a procedural DAG using hybrid retrieval that combines neural discovery via goal-level and step-level embedding indexes with symbolic DAG traversal. This supports requesting specific prerequisites, next steps, or alternative paths to reach a desired goal.
You can incrementally update procedural memory from new logs without catastrophic forgetting by using incremental EMA updates. This method continuously adjusts index vectors and probabilistic DAG edge statistics as new experiences arrive, preserving previously learned workflows.
Knowledge verification filters out incorrect workflows before DAG insertion by applying an explicit LLM-based scoring rubric to mined sequential patterns. This ensures only high-quality, verified procedural candidates are committed to the graph.
The best way to learn multi-step procedures from agentic logs is to apply sequential pattern mining to extract recurring workflows, verify them with LLM scoring, and index them in a dual-index procedural DAG for later neural and symbolic retrieval.
lespacemanEvaluate completed WebBench tasks as PASS, PARTIAL, or FAIL using LLM-based scoring on traces and extracted data.
Renzo-TognellaIntegrate symbolic graph reasoning with neural embeddings and LLMs for constraint-based retrieval.
CarbeneAIIndex codebases into concept-to-file JSON maps with semantic search.
KanakMalpaniA tagline describing your app.
