KNOWLEDGE MANAGEMENT FOR AGENTS

KEEPING KNOWLEDGE FRESH AND AVOIDING STALE INFORMATION Knowledge goes out of date. Prices change, policies are revised, products are discontinued, and people change jobs. Stale information is information that was once correct but is no longer, and it is dangerous because the agent states it with the same confidence as correct information. Nobody notices, because nothing visibly breaks. A knowledge base entry stating "our premium plan costs $29 per month" might remain in the system for years after the price rises, and the agent will quote the old price to customers. To fight staleness, good knowledge management uses several practices. Record the date. Every entry should carry a last-updated date and, ideally, an owner, meaning the person or team responsible for keeping it correct. Set review intervals by category. Different kinds of facts go stale at different speeds. Prices, promotions, and staffing lists change fast and might be reviewed every month or quarter. Policies and procedures may be reviewed every six months or annually. Slow-changing material, such as the company's history or glossary of terms, may be reviewed less often. Automatically flag overdue items. For instance, an entry about pricing is automatically flagged for human review after ninety days, so someone confirms or updates it. Link to the source of truth. Where possible, have the agent fetch live data, such as the current price from the pricing system, rather than relying on a copied value in a document. Re-index after changes. If you use RAG, updating a document means re-embedding and re-indexing it, or the search will still return the old version. Handle conflicts and retire old versions. When two documents disagree, the agent may pick either one. Remove or clearly mark old versions as superseded. Show provenance. Have the agent cite its sources and show the last-updated date, so users can judge freshness themselves. A final habit is to measure. Track how many entries are overdue, and test the agent periodically with questions whose answers recently changed, to confirm that it gives the updated answer.