“A Truth Maintenance System” was written by Jon Doyle at the MIT Artificial Intelligence Laboratory and issued in 1979 as AI Memo 521; it was also published that year in the journal Artificial Intelligence, volume 12, pages 231 to 272. The version cited here is the memo archived in MIT’s DSpace repository. A truth maintenance system, or TMS, is described in the paper as a problem-solver subsystem whose job is to record and maintain the reasons for a program’s beliefs.
The central idea is that a reasoning program should not just hold conclusions, it should remember why it holds them. Each belief in a TMS carries justifications, which are records of the other beliefs that support it. Because the support is tracked explicitly, the system can tell which beliefs currently rest on valid foundations and which have lost their support. When a contradiction appears, the TMS performs dependency-directed backtracking: instead of blindly undoing the most recent choice, it traces the chain of justifications back to the specific assumptions responsible for the conflict and revises exactly those.
This bookkeeping made it practical to build programs that reason with assumptions and then retract them gracefully, which connects the TMS directly to the nonmonotonic logics that McCarthy and Reiter were formalizing in the same period. The approach, and its later relative the assumption-based truth maintenance system, became a standard component of expert systems, configurators, and constraint solvers.
Why a business reader should care: any automated system that makes tentative decisions, discovers a conflict, and must back out cleanly, from a product configurator to a scheduling engine, is doing what Doyle’s truth maintenance system pioneered, namely keeping an auditable trail of why each conclusion was reached so that revisions stay consistent.