Approach (concept)

Think of agent memory like any stateful system: it needs history, review and restore points. The "care" language is a metaphor for that maintenance.

Lifecycle

1. Provenance. Record where every entry came from and when.
2. Detect. Flag entries using transparent signals: source trust, instruction-like language inside data, contradiction with earlier trusted entries, and the actions an entry could trigger. In our synthetic demo this helps surface the planted entries; it also produces a false positive and a miss, on purpose.
3. Quarantine. Hold flagged entries out of the agent's working context while a person or policy reviews them.
4. Recover. Restore to the last clean checkpoint and compare the agent's answers before and after.

How this relates to other tools (opinion)

Prompt-boundary filters, trace viewers and memory libraries address different layers. We think the memory lifecycle deserves attention of its own. This is our opinion; we have not verified how any specific product handles it.

Limits

Heuristics miss things and flag harmless things. Nothing here is a validated control, and no accuracy figure is offered. Status: concept / research preview.