Zep Agent Skills: Procedural Memory for Production Agents
Memory systems for production LLM agents mostly store semantic memory (facts about users and domains) or episodic memory (earlier interactions).
Research Systems Overview
- Trace2Skill
- WikiSkill
- SkillRevise
Research systems such as those listed above learn procedures from agent runs, package them as Skills, and evaluate them on benchmarks.
We describe Zep Agent Skills, a production service that gives deployed agents this procedural memory.
Evidence and Governance
The service operates with strict quality control:
- It uses only verified attempts as evidence.
- It cites the supporting events for each field of a Skill.
- It governs the admission and distribution of each Skill version.
- It records each task attempt as a Trace with ordered events, an outcome, and a verification strength.
Skill Compilation Process
A compiler with separate Reflector and Curator roles turns verified Traces into versioned Skills.
Admission and Distribution
Scans and admission gates control which candidates become active, and verified use outcomes change the search rank of each Skill.
Open Standard and Interoperability
The Agent Skills format is an open standard that Anthropic originated and donated to the Agentic AI Foundation, so any host of the format can load a compiled Skill.
Retrieval Mechanism
Agents retrieve Skills through lexical search with progressive disclosure, over an API or MCP.
Evaluation Benchmarks
We evaluate Zep Agent Skills on:
- $τ^3$-bench banking
- AppWorld
- EnterpriseOps-Gym
Performance Results
Skills compiled from the Traces of a teacher model increase the pass rate of three weaker executor models on the same tasks, at every reasoning effort, by 4.5 to 18.8 points.
Most of the gain is on the tasks that the teacher solved.
On ALFWorld, Skills compiled from the runs of a small model give a gain similar to or larger than the gain that prior work reports for continuous fine-tuning, with no training.