What it is
Agent memory is the mechanism that lets an agent store and retrieve information between sessions without holding everything in the context window. The agent accumulates knowledge about the project, decisions and preferences in external storage (files, a database, a vector index) and pulls in what's relevant on demand. It cures the amnesia: every new session used to start from zero.
Where it came from
- The 2024–25 pain: the agent forgot context between runs, and teams re-explained the project to it by hand.
- 2025 — Anthropic ships the Memory tool (storage in a memory directory; keeping what matters across compaction boundaries) plus context editing.
- An explosion of community solutions: claude-mem (★74.8k) — "the cure for session amnesia", agent-agnostic across Claude Code, Codex, Gemini CLI and Copilot.
Why it took off
- Long agentic tasks don't fit in the window: memory moves knowledge outside and pulls back only what's needed (benchmarks claim token savings in the tens of percent).
- Continuity: a new session sees where you left off and which decisions were made — far less repetition.
- It lined up with context engineering: memory is a natural part of managing agent state.
How to use it today
- Give your agent external project memory: key decisions, agreements, and "where to pick up from" (like STATE/JOURNAL in the Quest pipeline).
- Write decisions and facts into memory, not raw logs — otherwise it fills up with noise.
- At the start of a new session, read the memory first, then act.
What to watch out for
- Memory is "context with a history": stale or wrong facts land there too and get carried forward. Prune it.
- Persistent memory holding private data expands the leak surface (see the lethal trifecta).
- Community tools keep data in their own databases — check where exactly your context lives (supply chain and privacy).