Research note
Search is not memory
Information retrieval, as a field, has spent decades on a well-defined task: given a query, rank documents by relevance. It has become extraordinarily good at it.
The personal-memory case violates the premise, because the query is the missing part.
01 · Premise
The query is assumed, and that's the whole difficulty.
Web search works because you can usually articulate what you want, and because the index is enormous, so many documents satisfy any reasonable phrasing. Personal search inverts both conditions: you can't articulate it, and exactly one document satisfies you.
The consequence is unforgiving. A near miss on the web is still useful; a near miss in your own archive is a failure.
Web search has many right answers. Personal recall has one.
02 · Residue
What survives in memory isn't lexical.
Months later, what remains of an encounter is typically its gist, its appearance, its approximate time and its purpose, not its wording. Verbatim phrasing decays fast; the sense of a thing lasts much longer.
So the clue you carry into the search box is systematically the wrong type for the index you're querying.
- Gist survives
- What it was about, in your words, not the source's.
- Appearance survives
- Especially for anything visual.
- Temporal position survives
- Coarsely: the season, the project, the trip.
- Wording decays
- The one thing keyword search requires.
03 · Closing the gap
Accept the clue types people actually have.
Semantic search addresses the gist problem: describe the idea and matching happens on meaning rather than tokens. That's necessary but not sufficient, because it still expects a single well-formed description.
The rest closes by letting time, source, kind, appearance and adjacency act as query terms in their own right, and by letting the search be iterative. Retrieval then resembles reconstruction, which is what recall was all along.
Next thought
Reconstruction leans on relationships. Those relationships form a graph.
Read: the personal memory graphIf this idea interested you…
Shelf V3 is out now
Shelf V3 is the current build. Shelf does not use automatic updates, so download the latest build here and install it over your current version.
This is intentional. Shelf is built around a local-first, privacy-focused philosophy. An automatic updater requires additional background processes, network connections, update infrastructure, and another system with permission to modify the app. We prefer to keep Shelf’s network footprint and background activity as small as possible.
Manual updates keep the process explicit: Shelf only checks the internet when a feature actually needs it, and you decide when a new version is downloaded and installed.