Memory
Facts, preferences, decisions, and lessons Spinup keeps with the agent, above any one harness or runtime.
Memory is the durable record of what a Spinup Agent has learned. It holds facts, preferences, decisions, and lessons that should outlive any single run.
Memory lives with the agent, above any one harness or runtime. When the environment is replaced, when the harness is switched, when a runtime instance is stopped, memory stays.
How memory changes
You can edit memory directly through Spinup. A harness can also propose memory back to Spinup through the state tools it has available during a run.
Spinup treats a proposed change as a proposal, not a write. The proposed change carries the harness's claim about what should be remembered. Spinup decides whether to accept it.
Only after Spinup returns an accepted receipt is the memory saved. The harness should not assume it persisted before then. Anything the harness wrote to its own files is local to that environment and is not canonical until the receipt comes back.
What this gives you
Memory does not get trapped inside one harness's local store. The same facts and preferences are available whether the next run lands on OpenClaw or Hermes. The record of what was proposed, what was accepted, and what was rejected is recorded as evidence on the run that produced it.
Where to go next
- See Spinup Agents for what else lives with the agent.
- See Harnesses for the engines that propose changes back.
- See Runs for the unit that carries a proposed change.
- See Evidence for the record of what was projected, accepted, or rejected.