An agent that forgets everything between steps can't finish real work. But stuffing every past event into a prompt quickly becomes slow and expensive. Good agent memory is about deciding what to keep, where to keep it and when to bring it back.

Three layers of memory

Working memory

The current task, recent tool results and the plan. This lives in the model's context window and changes constantly.

Long-term knowledge

Facts, documents and past decisions that may be useful later. Store them in a vector index or a structured database and retrieve only what's relevant to the current step.

Durable state

Files, installed packages, open sessions and progress markers. On Codexa this lives in the sandbox itself and survives restarts through snapshots.

Practical tips

• Summarise completed sub-tasks instead of keeping raw logs.

• Tag memories with time and source so the agent can judge freshness.

• Write progress to disk often, so a crash never costs more than a few minutes.

• Review what your agent retrieves; most bad answers start with bad context.

With these three layers in place, agents can work on a task for days or weeks and still know exactly where they are.