Agents now carry lessons from one task to the next

Agents can now turn outcomes, corrections, and failed approaches from their task work into durable memory automatically. The next task starts with lessons from earlier work instead of relying on someone to copy them into the agent's notes. We built it because a correction you gave on one task was lost by the next unless someone copied it into the agent's notes.

Every substantive successful agent run refreshes the task's rolling summary. It preserves decisions, evidence, reviewer corrections, reusable lessons, unresolved work, and blockers while removing details that are resolved or superseded.

Task Machine periodically compiles the latest summaries into the agent's bounded memory. It reads each affected task's current summary once, not full transcripts or earlier summary revisions, so memory is based on the latest handoff rather than an ever-growing log. If the work produced no reusable lesson, the memory can stay unchanged.

Improve memory without another approval step

Automatic updates publish as immutable memory versions. Earlier versions remain available in history, so each change is inspectable and reversible. A manual Tidy memory remains separate: it proposes a consolidated version and waits for you to accept or discard it before anything changes.

Read Memories for how agents carry durable notes into future work.