To keep or not to keep? People continually face variations of this decision as they encounter information. A large percentage of information encountered is clearly useless — junk e–mail, for example. Another portion of encountered information can be "used up" and disposed of in a single read — the weather report or a sports score, for example. That leaves a great deal of information in a middle ground. The information might be useful somewhere at sometime in the future. Decisions concerning whether and how to keep this information are an essential part of personal information management. Bad decisions either way can be costly. Information not kept or not kept properly may be unavailable later when it is needed. But keeping too much information can also be costly. The wrong information competes for attention and may obscure information more appropriate to the current task. These are the logical costs of a signal detection task. From this perspective, one approach in tool support is to try to decrease the costs of a false positive (keeping useless information) and a miss (not keeping useful information). But this reduction in the costs of keeping mistakes is likely to be bounded by fundamental limitations in the human ability to remember and to attend. A second approach suggested by the theory of signal detectability is relatively less explored: Develop tools that decrease the likelihood that "keeping" mistakes are made in the first place.