An organisation's coding agents forget between sessions, and what one repository learned does not reach the next unless something carries it. These are working notes on a memory shared across repositories, the organisation above them and the models of several vendors — on what a read must do for that knowledge to arrive, and on what it cannot do. They are for someone deciding what to demand of a shared memory for agents, not someone deciding whether to buy one.
An approximate index that filters after it has chosen, and a history filed under a key that changed form: two ways a memory holding the answer returns nothing, one reproducible in a short SQL script against a public container image.
Paper 002Ranked in one read by nearness of scope, a large repository takes every candidate and the organisation's rule is never considered. A second read, the reason its scores compare, and a bar that stops the seat being guaranteed.
Paper 003Distillation restates one fact many times. Overlap measured against the larger sentence, a cap of two blocks per source file, and the rule that a disagreement is never a repeat — with the reversal it still misses.
Paper 004One store behind an open protocol serves agents on several vendors' models. What the harness still decides, what a credential cannot attest, and the open items, each with what would settle it.
A memory shared by an organisation's agents is a second layer only if the organisation's knowledge survives the read, and each way it fails to is silent. An approximate index filters after it has chosen: at pgvector 0.8.4's defaults, seven or eight of twenty scoped reads in Paper 001's public experiment returned nothing from a scope holding 2,000 rows, and none did with an iterative scan. Ranked in one read by nearness of scope, a large repository can take every candidate, so the organisation needs its own read, scored on the same footing; a fixed budget needs overlap measured symmetrically, and disagreement never counted as repetition. None of this settles whether an agent consults the memory, which model read it, or whether a sentence in it is still true.
The methods here were developed inside SkilledMind, a Digital One property, and the defects described here were found while testing it, by queries against it and by its own test suite. That repository is not public, so its behaviour is its builders' account rather than something a reader can check today. The index experiment in Paper 001 runs against a public container image, and every overlap ratio in Paper 003 can be recomputed by hand. Vendors are named where their documentation is quoted or where their tools were configured against the memory. Deliberately absent: prices, plans and tiers, and deployment internals.
These notes are contributed under CC BY 4.0: quote, reproduce and build on them with attribution, a link to the licence, and adaptations marked as such. The Foundation is an entity in formation; no public-benefit (ANBI) status is claimed at this time.