Lemar is no longer just reactive. It now works with persistent goals, checks whether actions produced the expected outcome and corrects the plan when something fails.
What changed
- Persistent goals. Every relevant case now has a goal with priority, deadline and progress. Tasks are recorded and survive between runs.
- Outcome verification. After acting, Lemar schedules a real check: the repair changed status, the order was sent, the refund was issued, the customer received the message. If it cannot confirm, it retries and replans.
- Operational memory. It stores what worked and what failed by action, supplier, customer and repair. That memory feeds the next planning step.
- Replanning with limits. When a step fails, Lemar creates new tasks, but with limits on attempts, pending tasks and budget.
- Budgets per goal. Every goal has limits for cost, time, model calls and attempts. When the limit is reached, it stops and asks for review.
- Sandboxed tools. Actions are classified as read, network, write and money. Permissions, audit and limits still apply.
- Self-evaluation. Every day it compares the last 7 days with the previous 7 and warns if completion drops or failures rise.
- Reviewed learning. Gaps still become candidates for review. Nothing is published automatically and Lemar does not modify its own code.
How to use it
- Follow the unconfirmed-outcome and regression notices in the Autonomy tab.
- When a goal fails repeatedly, review the action, supplier or training knowledge.
- Keep value limits and budgets aligned with the trust you have in each flow.
Lemar still proposes before risky actions and respects the shop limits. The difference is that it now closes the loop: act, observe, learn and correct.