The organisation needs its own read.

A memory shared by an organisation's repositories is only a second layer if what the organisation knows reaches the agent working in one of them. Ranking both in a single read by nearness of scope works against that in a large repository: the scope term hands the repository a lead on every candidate, and in the case measured here every seat went to the repository while the organisation's rule for the task was not a candidate. The repair is a second read, and the reason its scores can be compared with the first.

§1Nearness of scope is a sound ranking rule, and a greedy one

The memory these notes describe files everything under one of three nested scopes: a repository, a team, the organisation. When an agent in a repository asks for a brief, the read spans that repository and the organisation, and every candidate is scored on six terms: similarity of meaning, shared words, recency, the standing of the layer it was distilled into, which part of the repository it came from, and scope proximity — how near its scope is to the one the agent is working in. The weights are fixed:

score = 0.55 meaning + 0.17 words + 0.06 recency + 0.05 layer + 0.09 scope + 0.08 path

scope proximity, for a read of [repository, organisation]
  the repository     1.0
  the organisation   0.0        a team, when named, sits between at 0.5

The rule is right in its own terms. A note written in this repository is more likely to be true here than the same sentence written elsewhere, and when two memories say similar things the nearer one should win. But 0.09 is a constant offset, and a constant offset does not compare two memories — it compares two populations. To outrank a repository memory, an organisation memory must lead it by 0.09 on the other five terms combined; to enter a list of 32 it must lead the 32nd-best repository memory by that much. In a repository with fewer memories than the list holds, the organisation enters because there is room. In one with a long history it needs a lead that the prompt tested here did not give it.

§2The first seat was offered inside a list the organisation could not enter

The first repair gave the organisation a seat. When a read spans a repository and the organisation, the two best organisation hits within 85% of the top score are moved to second and fourth place in the brief, so that what the organisation knows is carried even when a repository has a great deal to say. The candidate list was raised from 16 to 32 at the same time, to give the seat room.

It was tested the way it would be used: a brief requested from inside a repository with a long history, for a task squarely inside the organisation's own rules.

prompt: "write a new Foundation paper about shared agent memory"
read over [repository, organisation], k = 32
  candidates from the repository     32 of 32      score 0.578 … 0.539
  candidates from the organisation    0 of 32

the same prompt, read over [organisation] alone, k = 5
  0.541  what the organisation's shared memory is
  0.495  papers here make one claim, in 5 to 8 minutes       the rule that governs the task
  0.485  ground generations in retrieved context
  0.459  the default model and harness for coding agents
  0.445  an unrelated line

the bar, 0.85 × 0.578 = 0.491   → two organisation memories clear it

The arithmetic of §1 explains it exactly. Scored inside the combined read, the organisation's best memory loses the 0.09 it would earn as the nearest scope and lands near 0.45, below the 32nd repository candidate at 0.539. The seat looked for organisation hits among candidates that the scope term had already made sure contained none. A larger list only moves the problem: it admits the organisation once it reaches below about 0.45 in repository scores, and how deep that is depends on how much the repository holds, not on the question asked.

§3A second read puts the organisation on the repository's footing

The repair runs a second read beside the first whenever a request spans a repository and the organisation: the same prompt, restricted to the organisation's scope alone, four candidates deep. The two run concurrently, so the second adds a query to the database's work but not, on its own, to the wait; its latency was not measured.

The comparison is what needs arguing. Scores from two different reads are not in general comparable, and a bar of 85% of the top score is meaningless if the two tops were computed differently. Here they are computed the same way: every term is a function of the query and the memory alone, and none is scaled against the other results of its read. In a read of the organisation alone, the organisation is the nearest scope, so its best memory earns the full proximity term — exactly as the repository's best memory does in the combined read. Both tops sit on the same footing, and an organisation memory is held to the bar it would have faced had it been written in the repository itself. The best two across both pools that clear it take the second and fourth places — and an organisation memory that also made the combined list is judged by its score from the organisation read, never by the lower one it carries there, so that reaching the list can never make a memory harder to seat.

After the repair was deployed the same prompt produced a brief with organisation memories second and fourth, at 0.541 and 0.495; the other two the organisation read returned, at 0.485 and 0.459, fell below the bar of 0.491 and were left out. A second prompt — how a new paper here should be structured and how long it should be — put the organisation's rule second; the next organisation hit, at 0.507, fell short of the bar of 0.85 × 0.654 = 0.556, so the fourth place stayed with the repository. That is the intended shape: the bar is what stops the seat being guaranteed. A test fixes the property in place: forty repository memories that alone fill the candidate list, one organisation memory on the same topic, and an assertion that it reaches the first two places of the brief. It fails against the first repair and passes against the second.

§4What a seat cannot judge

The seat decides where the organisation speaks. It has no view on whether what it says is worth hearing. Four limits remain.

A seat is taken by whatever the organisation holds. Text filed at the organisation's level that nobody in the organisation wrote — sample content, a bulk import of reference material — competes for the seat on the same terms as a decision somebody stated, and on a prompt where its wording happens to be closer, it wins. What reaches an agent through this device is exactly as good as what is filed at that level, which makes curation of the organisation's scope a duty rather than a nicety.

The bar is relative. Eighty-five percent of a weak top is a low bar. When the repository has nothing close to the prompt, the organisation's seat is easier to win, which is usually right and occasionally admits a line that is merely less irrelevant than the rest. The figure 0.85 is a chosen setting, not a measured optimum, and nothing here claims otherwise. Nor is the margin wide: on the first prompt the rule cleared the bar by 0.004, so its seat there rests on the third decimal place, and a slightly different wording could lose it.

A read cannot find what was never written. The organisation's strategy, its positioning, the decisions it has taken about how it works — these reach an agent only if a person has stated them at the organisation's level. Distillation from repositories produces what repositories say; it does not produce what the organisation decided. For an organisation that wants its agents to know its strategy, the first step is not in the memory at all.

The second read is the organisation's alone. A team, when named, gives up 0.045 on every candidate and has no read or seat of its own; whether it is crowded out the same way was not tested.

How repeated practices are promoted from repositories to the organisation, and what a new project may then be told on its first day, is the subject of The Inherited Default; this paper assumes that stage and is concerned only with what is read back.