Give Your Agent a Memory

An agent without memory makes you its context: you re-explain preferences, decisions, and house rules every session. An idapt agent remembers, and (the part that matters) you can read exactly what.

How it works

Every agent has its own memory: structured notes it writes and recalls. Tell it "always cite primary sources" or "the deploy target is staging first" and it records the rule; next session, in any chat, the rule is simply in force. The agent recalls relevant entries when they matter rather than dumping everything into every prompt.

Memory is legible by design: open the agent's memory and you see the entries as text. Edit them, delete them, add your own. No black box deciding who you are.

Memory, context, files: three lifetimes

Confusing these causes the classic failures, so keep them straight:

Context is this conversation. It resets.
Memory is what the agent carries across conversations: preferences, decisions, standing facts. Small and curated.
Drive is the work itself: documents, data, artifacts. Durable and big.

The rule of thumb: decisions and preferences go to memory; deliverables go to files; everything else can stay in context and die there. Stuffing documents into memory bloats recall; leaving decisions in context means re-litigating them next week.

What good memory habits look like

Seed it at creation: give a new agent five memory entries about its job and your standards; the first run behaves like the tenth.
Correct it out loud: when the agent misjudges, say "remember: X, not Y" and check the entry it writes.
Audit monthly: prune entries that no longer hold. Stale memory is worse than none because it fails confidently.

Good to know

Memory is per-agent, not shared: your careful research agent's rules never leak into the fast drafting one.
The same memory serves the agent everywhere it runs: chat, API, CLI, MCP.
What is agent memory covers the concept; the help article covers the controls.

Create one specialist agent today, seed five memory entries, and give it a real task: the difference from a bare model is immediate. The agents feature page has the full anatomy.