Agent memory
Knowledge an AI agent keeps across sessions, project decisions, house rules, past mistakes, so each run does not start from zero.
By default an agent session starts with amnesia. The model knows its training data and whatever is in the current context. It knows nothing about last week's decisions. Agent memory is the layer that fixes this. It keeps durable records of what was decided, what failed, and what the project's ground truths are, and feeds them into future sessions.
Implementations range widely. Some are plain files in the repo: memory documents, learnings files, AGENTS.md-style rules. Others are retrieval systems that surface relevant past episodes on demand. Across all of them, one finding repeats. Decisions and corrections are the highest-value content. Knowing what was tried, rejected, and why prevents the same wrong turn twice.
Memory has an admission problem as well as a storage problem. A store that keeps every observation drowns the useful entries. The systems that work curate hard. They also record where each entry came from, so a stale memory can be traced and retired.