Isomorph
← Back to console

Agent state: code and memory

Every agent is three layers: a fixed core that never changes, the code it runs (its skills and capabilities), and the memory it has learned. Code and memory are saved and switched independently. The description of what the code can do always travels with the code, so a switch can never leave the agent “remembering” a skill it no longer has, or missing one it does.

Code: what the agent runs

The code is the versioned body of the agent: the software it runs, its skills, and the always-true description of what this code can do. That description ships with the code, so it is always in sync with it.

  • Standard starting points are the stock versions we provide. They are read-only baselines.
  • Saved code versions are snapshots you capture from a running agent, tagged with a name, the exact code revision, and the capture time.
  • Deploying a code version rebuilds the agent from that code. Its learned memory is kept. A full backup is taken first.

Memory: what the agent remembers

The memory is everything the agent has learned: what happened, its notes, and what it knows about you and your business. It does not include the description of what the code can do, because that lives with the code, so switching memory never changes what the agent can do.

  • Fresh / empty wipes the learned memory to a clean slate. Built-in capabilities are kept. (Shown with a warning, because learned memory is cleared.)
  • Saved memory versions are snapshots you capture, tagged with a name, how many memories they hold, and the capture time.
  • Restoring a memory version replaces the learned memory with that snapshot. Code and capabilities are unchanged. A full backup is taken first.

What the agent keeps when it updates

Memory is split by where it lives, which guarantees capabilities and memory never drift apart:

  • Core memory (fixed): what the agent is, how to use it, built-in troubleshooting. Identical in every code version.
  • Capability memory (ships with the code): what this code can do. Switching code refreshes it to match the new code.
  • Learned memory: what the agent picked up while working, its decisions and notes. The only part a memory restore touches.

Saving, restoring & backups

  • Save current state captures the live code and/or memory as a named restore point: save both together or just one.
  • Every switch is backed up first: the prior state is tagged automatically and a full snapshot of the agent's disk is taken, so any change is reversible.
  • Saved versions and snapshots are bounded by per-agent limits and an automatic retention window, so they do not accumulate without limit.

Good to know

  • Switch and save actions are queued and applied within a few minutes. Refresh the State tab to see the result.
  • Deploying a code version briefly restarts the agent; a memory restore lets running work finish first.

Manage these from an agent's Agent State tab in the console.