AGENT ATLASa field guide to agent workflows
Back to Memory & context
Memory & context

Long-term memory

Remember across sessions — facts, preferences, past work.

N agentsAdvanced

How it works

  1. 1Decide what's worth persisting — durable facts, not transient chatter.
  2. 2Write those to a store, keyed for retrieval.
  3. 3On a new session, retrieve relevant memories into context.
  4. 4Periodically prune or update stale or contradicted memories.

Use it when

Assistants that should recall user preferences, prior decisions, or accumulated knowledge between separate conversations.

Reach for something else when

One-off tasks. Persistent memory adds storage, retrieval, and staleness/privacy concerns.

Where you stay in the loop

You decide what is worth remembering and what should be forgotten; the agent persists and recalls. Review stored memories for staleness and privacy — durable memory is a responsibility, not just a feature.

In the wild

A coding assistant remembers your stack and conventions so you don't restate them each session.

Hand this to your agent

Act with long-term memory.

At the END of each task, output a MEMORY block of durable facts worth keeping
about me or the project (preferences, decisions, conventions) — concise, no
transient details.

At the START, I'll paste the saved MEMORY; use it as context and tell me
which memories you applied. Flag anything that now looks outdated.

First task: <...>

Replace the <…> placeholders, paste it into your agent, and it'll scaffold the workflow with you.