Creativity · Agent Protocol
Agent Procedural Memory Pattern
Procedural memory is the 'skill library' — not facts about the world, but learned procedures: 'to deploy to staging, run these three commands', 'when the API returns 429 for longer than 60s, escalate to on-call'. Agents with procedural memory improve over time on repeat tasks rather than starting each session from scratch.
Protocol facts
- Sponsor
- Cognitive architecture research
- Status
- proposed
- Interop with
- Voyager, LangGraph, AutoGen, skill libraries
Frequently asked questions
How does an agent learn a procedure?
After completing a task, the agent summarizes what worked into a named skill ('deploy_staging', 'fix_flaky_test') and stores it. On future similar tasks, retrieval surfaces matching skills as hints to the planner.
Who popularized this pattern?
The Voyager paper (NVIDIA + Caltech, 2023) demonstrated a Minecraft agent that built a growing skill library, unlocking increasingly complex behaviors without any RL fine-tuning.
What's the risk?
Brittle skills: a skill learned in one context fails in another. Mitigate with preconditions (when the skill is applicable), post-hoc validation, and versioning so you can roll back bad skills.
Sources
- Voyager paper — accessed 2026-04-20
- CAMEL — role-playing agents — accessed 2026-04-20