Most guides to building an external memory start with structure: a folder scheme, a tag taxonomy, a template. That order is backwards, because structure is the part you can automate and habit is the part you cannot. This is a four-week plan that inverts it—capture first, let software handle classification, and run a concrete retrieval test on day 30 before you invest any more effort.
Week 1: one stream, no filing
Pick a single place for everything you save and use it for text, voice, and photos alike. The rule for the week is that nothing gets filed, titled, or tagged at capture time. You are measuring one thing: whether the capture cost is low enough that you do it while distracted.
The target to beat is three actions from a locked phone to a saved thought. Above that, capture fails during exactly the moments worth capturing—walking, queuing, leaving a call. If you end week one with fewer than fifteen items, the problem is the capture path, not your discipline, and no amount of structure downstream will fix it.
Week 2: capture the four things people skip
Most archives are full of articles and empty of the material that actually gets searched for later. Deliberately add these four for a week:
- Decisions with their reason—‘we chose the annual plan because the monthly one broke the budget in Q3’—not just the outcome.
- Commitments you made out loud, with the person's name and the rough deadline attached.
- A thirty-second voice debrief after any conversation that mattered, recorded while walking away from it.
- Photos with one line of explanation: the whiteboard, the label, the receipt, the parking level.
Week 3: let the software do the organizing
Now check what is being derived automatically: transcripts for voice, extracted text for photos, suggested topics, and tasks pulled out of ordinary sentences. Correct the ones that are wrong—this is maintenance worth doing, because it is bounded and takes seconds.
Do not build a taxonomy. Every hour spent designing categories is an hour spent predicting questions you have not been asked yet, and the prediction is usually wrong. If a genuine cluster emerges—one client, one recurring theme—name it then, from evidence rather than in advance.
Week 4: practice retrieval deliberately
Retrieval is a skill and it degrades unused. Three times this week, ask your archive something you could answer from memory, and compare. You are calibrating: learning which phrasings work, and where the archive is thin.
Ask by description rather than keyword—‘the pricing conversation where someone objected to the middle tier’ rather than ‘pricing’. Add a time filter when you remember roughly when it happened. If an answer is assembled for you, open at least one source note behind it every time, until you have a reliable sense of when the summary can be trusted.
The day-30 test
Write five specific questions about the last month before you search: What did I promise Marta? Why did we drop the middle tier? What was the tool from the Thursday call? Where did I park at the airport? What did I decide about the contractor? Then answer them from the archive alone.
Three out of five in under a minute each means the system is real and worth continuing. One or two means capture is too narrow—you are saving reference material and skipping decisions and commitments. Zero means the capture path is broken, and the fix is at the beginning of the pipeline, never at the end.
What this plan cannot do for you
It cannot recover anything you did not capture, and it will not feel useful during weeks one and two. That gap is structural: for the first few weeks the archive is small enough that ordinary memory beats searching it, so the work is pure cost with a delayed return. Most abandonment happens precisely here, before the payoff exists.
Two other limits matter. Automatic topics are probabilistic and will mislabel some material, so treat them as suggestions rather than guarantees. Short, context-free notes—‘call him’, ‘check this’—remain hard to retrieve because there is almost nothing to match against.
Do not design the system before you have the habit. Capture into one stream for thirty days, let software handle topics and transcripts, then run the five-question test. Three correct answers earns the next month of effort; fewer means fix capture, not structure.