# Tuning Memories Fade Korium never hides a memory because it is old. Age lowers a memory's rank in search instead, and one setting decides by how much. This guide explains the setting, shows the arithmetic with worked examples, and gives you a way to test a change before you keep it. ## Why memories fade and are never hidden {#why} Korium used to hide a memory once it reached a set age for its type. A correction was gone from search after 90 days, an insight after 180, a lesson after 365. That took true memories with it. If an agent saved a correction and nobody replaced it, the correction disappeared after 90 days, the agent made the same mistake again, and nobody could see why. The one useful thing that hiding did was stop old memories from crowding new ones out of a page of results. That is a question of order, not of what exists. So Korium now handles it as order: a memory stays current until something replaces it, and its age only moves it down the list. A memory that fades is still there. A search that matches it can return it, you can browse to it, and a replacement link still points at it. Fading decides which of two memories that match about as well comes first. ## Where the setting is and who can change it {#where} Memories Fade is set once for your whole organization. Every workspace in the organization follows it, and no workspace has a setting of its own. Open the Settings page of your organization's own workspace, the one at the top of the organization, and find the Memories Fade card. It holds two things: - **The Memories Fade switch.** It is on unless your organization's admin turned it off. Off means age has no effect on rank at all. - **The strength.** A number from 0 to 1. It starts at 0.2. When the switch is off, the strength is kept and not used, so turning the switch back on restores what you had. Only an admin of your organization can change either one, and only there. An admin of a workspace inside the organization who is not an admin of the organization cannot change it. Everyone else, on the organization's page and on every workspace inside it, sees the organization's values and who sets them. A change reaches searches in every workspace of the organization within a minute; nothing is restarted and no memory is rewritten. A strength outside 0 to 1 is refused, and nothing is saved. The best possible match is worth 1 in the score, so a strength above 1 would let age outweigh a perfect match. ## Why the strength is yours to set {#yours} The right strength depends on your data and on what you use Korium for. Korium cannot know either one in advance. A team that saves many short notes every day has a lot of recent material that overlaps. For that team a higher strength keeps this week's note ahead of the similar note from last spring. A team that saves a few careful decisions and lessons has little overlap, and most of what it saved stays right for a long time. For that team a lower strength, or the switch off, keeps a well matched older memory at the top. 0.2 is a starting point that suits a mix of both. It is small enough that a clearly better match still wins over a newer one, and large enough to settle a near tie in favor of the newer memory. ## How fast each type fades {#speeds} A memory saved today is fully fresh: its freshness is 1. Freshness then halves each time a set number of days passes. The number depends on the memory's type.
Memory typeFreshness halves everyFreshness after one yearWhat it has lost after one year, at 0.2
Drift, correction90 daysAbout 0.06About 0.19
Insight180 daysAbout 0.25About 0.15
Lesson, pattern365 days0.50.1
Decision, preference, rejection, reflection, handoff, seedNever fades10
Six types never fade. A decision, a preference or a rejection is a position your team took, and it holds until a new one replaces it, however long ago that was. A reflection, a handoff and a seed are kept the same way. A decision saved three years ago ranks exactly as it would if it were saved today. These speeds are the same for every workspace. The strength is the part your organization sets. ## Which date the age counts from {#date} A memory's age counts from the latest of four dates: the day it was first saved, the day it became current, the day its label was last set, and the day it was last saved again. Usually all four are the same day. They differ when a memory is saved again. Saving a memory again makes it fresh from the day of the newer save. That holds for a short memory and for a long one such as a transcript, and whichever label the newer save carries. The date the memory shows does not move. A long memory is stored turn by turn, and you can save part of it again: the first 12 turns of a 20-turn transcript, with nothing new in them. Two rules decide which turns are then fresh, in this order. 1. **The newer label stays: the whole memory is fresh.** When the label of the newer save is the one that stays, every turn of the stored memory takes that label and is fresh from the newer save. The same label saved again counts. Take a lesson of 20 turns and save its first 12 turns again today as a lesson: all 20 turns are fresh from today. 2. **The stored label stays: only the turns you saved again are fresh.** Take the same lesson of 20 turns and save its first 12 turns again today as a drift note, which a lesson outranks. Turns 1 to 12 are fresh from today. Turns 13 to 20 keep the age of the first save. [How memory works](how-memory-works.html#supersedes) says which of two labels stays. ## The score, in plain words {#formula} Every result gets a score, and the page is ordered by it, highest first. ``` final_score = match + 0.6 × importance - strength × (1 - freshness) ``` - **Match** is how well the memory fits your question, scaled so that the best match in this search is 1 and the weakest is 0. - **Importance** is the number from 0 to 1 saved with the memory, or its type's default when none was given. - **Freshness** is the number from the table above: 1 on the day the memory was saved, then halving. - **Strength** is your setting. With the switch off it counts as 0. The last line is what a memory loses for its age. It is 0 for a memory saved today and for a type that never fades, and it grows towards the strength as the memory ages. So Memories Fade never raises a score: it only takes something off the old ones. Each result carries two fields so you can see this for yourself. `final_score` is the total. `fade_part` is the last line of the sum: 0 or a negative number, never more than the strength below 0. The older field `fused_score` is the match before it is scaled, and it does not change when you change the strength. ## What turning it on adds beyond reordering {#deeper} Fading can let a fresher memory from just below the page take the place of an old one near the bottom of it. Nothing else changes. With Memories Fade off, each half of the search (by words and by meaning) keeps one page of its best matches, and the results are ordered from those. With it on, each half keeps three pages, so that a fresh memory is not kept off the page by many old ones that match a little better. A memory from below the first page never passes one that is on it unless that one has faded more than it has. Its own match does not count. Before its own fade, it scores the lower of two numbers: - the lowest score, before fading, among the memories the first pages hold; - what its own importance earns with the weakest match those pages hold. On most pages the weakest match counts as 0, so this is 0.6 times its importance. When every memory the first pages hold matches equally well, as on a page of one, each match counts as 1, and this is 1 plus 0.6 times its importance. Then it loses its own fade. The first number keeps it under everything on the page that has not faded. The second keeps an unimportant memory from being scored as if it matched as well as a page result that matched well. So a fresh memory from below the page moves up only past results that have faded below its score. At the default strength of 0.2 a result loses at most 0.2. Importance still counts. A less important fresh memory needs a bigger age gap to pass a more important old one. Last week's drift note, with the default importance of 0.4, does not come onto a page of two-year-old lessons at 0.7: it gives up 0.18 in importance, and at a strength of 0.2 the lessons have lost only 0.15. At a higher strength it does. For example, 25 drift notes from two years ago all match a question a little better than last week's correction, and a page is ten results. With Memories Fade off the page is ten old notes. With it on at 0.2, each old note has lost almost 0.2, the correction has lost almost nothing, and the correction takes a place near the bottom of the page. If nothing on the page is old, or everything on it is of a type that never fades, the page is exactly the one you get with Memories Fade off. The last place on the page is where this shows most. When the last result is both the lowest-scoring memory the first pages hold and their weakest match, and it has faded at all, a fresh or never-fading memory from below the page that is at least as important takes its place. That can happen at any strength above 0: at 0 nothing moves. A result that scores above the lowest score on those pages has to fade by more than that lead before a memory from below can pass it. Among the memories below the page, the higher score comes first, and the better match settles a tie. ## Worked examples {#examples} Take three corrections that match your question equally, each with the default importance for a correction, 0.6. One was saved today, one 90 days ago and one 180 days ago, so their freshness is 1, 0.5 and 0.25. Match is 1 for all three and importance adds 0.36.
StrengthSaved todaySaved 90 days agoSaved 180 days agoWhat decides the order
01.361.361.36Nothing about age. The three tie.
0.21.361.261.21The newest comes first, by a small margin.
0.81.360.960.76The newest comes first, by a wide margin.
The memory saved today scores the same at every strength. Only the older two lose anything. Now take two corrections that do not match equally. The older one fits the question best: match 1, saved 180 days ago, freshness 0.25. The newer one is a weaker fit: match 0.6, saved today, freshness 1. Importance adds 0.36 to both.
StrengthOlder, better matchNewer, weaker matchWhich comes first
01.360.96The better match.
0.21.210.96The better match.
0.80.760.96The newer one. Age now outweighs the better match.
That last row is what a strength set too high looks like. A faded memory is lowered against everything, not only against newer memories of its own type. Take a correction saved two years ago that matches your question fairly well: match 0.43, importance 0.5. Beside it is a decision that is linked to another result and did not match your words at all: match 0, importance 0.9. A decision never fades.
Memories FadeThe old correction that matchesThe linked decision that does notWhich comes first
Off0.730.54The correction.
On, at 0.20.530.54The decision, narrowly.
The decision's score did not move. The correction lost 0.2, almost the whole strength, because after two years it has faded nearly all the way. If that is not what you want for your data, lower the strength or turn the switch off. ## How to tune it, step by step {#tune} 1. **Choose your questions.** Write down five to ten questions your team really asks, where you know which memory should come first. 2. **Record what happens now.** Ask each question. For the top results, note the memory, its type, the date it was saved, its `final_score` and its `fade_part`. 3. **Change the strength by a small step.** Move it by 0.1 and save. An admin of your organization does this on the organization's own Settings page, and the change applies to every workspace in it. 4. **Ask the same questions again.** Compare the order and the save dates with what you recorded. `fade_part` shows how much of each score your change moved. 5. **Keep it or put it back.** Keep the change if the memories you expected moved up and no better, older match left the first page. If you are not sure, put the old value back. Nothing was changed in the memories themselves, so going back is exact. Change one thing at a time. If you also change which memories exist, or how the questions are worded, you cannot tell what moved the results. ## Signs it is set too high or too low {#signs} **Too high.** Today's notes crowd out better older matches. You ask a question you know an older memory answers well, and the first results are recent notes that only touch on it. Their `fade_part` is close to 0, and the older memory's is close to the strength below 0. Lower the strength. **Too low.** Old notes crowd out new ones. Several notes on the same subject come back, and an old drift note or correction sits above the newer one that updates it. The `fade_part` values are all close to 0, so age is settling nothing. Raise the strength. **About right.** Among results that match about as well, the newer one comes first. A clearly better match still wins, whatever its age. If old and new memories disagree, do not use the strength to settle it. Save a correction and link it to the memory it replaces, as [How to find and correct](working-with-memory.html) describes. Fading orders memories; it does not decide which one is true. ## What it does not do {#limits} - **It deletes nothing and retires nothing.** At any strength, no memory stops being current because of its age. A faded memory can be pushed off the first page by fresher ones that match about as well; a more exact question or a larger result count brings it back. - **It learns nothing from use.** How often a memory was returned, opened or acted on does not change its rank. - **It does not change what a memory matches.** It changes the order, and with it on Korium looks further down each half of the search, as described above. - **It does not touch the six types that never fade.** Their freshness is always 1, so they lose nothing at any strength. - **It does not bring back memories hidden earlier.** A memory that was marked no longer current under the old expiry rules stays that way. - **It does not replace a correction.** A memory that is no longer true needs a replacement, not a lower rank. [How memory works](how-memory-works.html#relevance) describes the whole of ranking, and the [tool reference](tool-reference.html#search-response) lists every field a search returns.