agent-memory-patterns
Installation
SKILL.md
Agent Memory Patterns
Overview
Memory is the foundation of agent intelligence. Without structured memory, agents repeat mistakes, lose context across sessions, and cannot build compound knowledge over time. A well-designed memory system determines what the agent remembers, how it retrieves relevant context, and when it safely forgets.
When to use: Designing a stateful AI agent or autonomous workflow; reviewing agent code for context management; evaluating retrieval latency or cost; any system where an agent must persist knowledge across conversation turns, sessions, or agent boundaries.
Quick Reference
| Memory Type | Storage | Retrieval Strategy | Lifetime | Use Case |
|---|---|---|---|---|
| Short-term | Context window (in-memory) | Direct inclusion — no retrieval needed | Single session | Active conversation, current task state |
| Working | Scratchpad / todo file | Sequential read — agent writes and reads directly | Task duration | Reasoning steps, partial results, sub-goals |
| Long-term episodic | Vector DB with timestamps | Embedding similarity + recency weighting | Months to permanent | Past interactions, specific sessions, event log |
| Long-term semantic | Structured store / knowledge graph | Keyword or concept-graph traversal | Permanent until invalidated | Facts, entities, domain knowledge |
| Procedural | File store / instinct YAML | Template match on task type | Permanent | Reusable patterns, learned workflows, instincts |