Scheduled Agents

Scheduled and ephemeral agents need more than a diary from the previous run. ZMem checks whether prior state is current, stale, or unknown before that state reaches the agent.

Run A Scheduled Agent

zmem scheduled-run \
  --session-id cron://daily-signal \
  --agent hermes \
  --task "collect the daily product signal" \
  --scope project:zmem \
  --stale-after-seconds 86400 \
  --summary-only \
  -- your-agent-command

The command writes a cold-start audit, starts the session, injects admissible memory, runs the command, checkpoints the result, and links those records in one proof digest.

Use --handoff-dir .zerker/handoff or --snapshot path/to/snapshot.json to restore into an empty database before the audit.

Continuity States

• current: the latest completed lifecycle state is within the configured wall-clock gap.
• stale: the latest completed state is older than the configured gap.
• unknown: there is no prior state, the previous session was not checkpointed, the latest snapshot payload is unavailable, or the clocks cannot establish ordering.

Stale or unknown does not mean every memory is false. It means the agent must not silently treat continuity as established.

Record Silent Success

zmem failure record \
  --expected "transfer exactly 10 credits" \
  --observed "API returned 200 but transferred 100 credits" \
  --correction "compare requested and settled ledger amounts" \
  --invalidation "settled amount differs from the request" \
  --confidence 0.98 \
  --scope project:payments \
  --agent payments-agent \
  --summary-only

The result is a typed episodic memory with a write receipt. Agent-authored failure memories enter quarantine. The correction becomes inspectable evidence, not an automatically trusted instruction.

Inspect it later:

zmem failure show <memory-id> --summary-only