Without a log you cannot apply a bonferroni-correction, estimate a false-discovery-rate, or answer the single most important question about any backtest: how many other versions did you try before this one.
Record the hypothesis before running the test, then the result and the decision. Writing the hypothesis first is uncomfortable precisely because it prevents retrofitting an explanation to whatever the data produced, which is p-hacking in its most natural form.
A log also compounds. Ideas that failed for a specific reason, wrong instrument, wrong horizon, too costly, are often revivable later under different conditions, and a year of notes is a far better starting point than a year of memories.
Related: multiple-testing, data-snooping, p-hacking, bonferroni-correction