Agent memory is the durable state an AI agent can carry across sessions: useful facts, decisions, preferences, constraints, and project knowledge that should influence future work. It is different from a context window. Context is what a model can see now; memory is what the system has decided should remain useful later.

A practical agent-memory system needs more than storage. It needs a lifecycle: capture evidence, decide what deserves to persist, assign scope and lifetime, retrieve only what is relevant, preserve provenance, and replace or forget state when reality changes. A transcript records what happened. Memory is a maintained judgment about what the next agent should know.

After more than 1000 coding-agent sessions across Claude Code, Codex, Cursor, Pi, and OpenCode, I found the stable architectural boundary is user-owned state beneath replaceable models and harnesses. The detailed evidence, architecture, and failure modes are documented in my Agent Memory field report.