An AI memory proposal would preserve organizational judgments—not just records

Finding the documents behind a decision is not the same as recovering why that decision once made sense. A proposed “shared organizational cortex” aims to preserve that reasoning—and track when its foundations no longer hold.

Organizations accumulate emails, reports and meeting records, but their decisions depend on interpretations assembled from those materials. A partnership might look promising because several separate observations align. Recovering the observations later does not necessarily recover the judgment, its assumptions or the circumstances that could undermine it.

In a preprint, Andrew Miracle, Phyl Georgiou, Christine Okoth and Ese Monday propose a “shared organizational cortex” to retain those interpretations. The phrase describes a knowledge architecture, not a demonstrated artificial brain. Its intended capability—what the authors call organizational instinct—remains a design objective rather than an evaluated result.

Four layers, with different jobs

The proposal starts with immutable source records called Entries, then organizes knowledge into four layers.

Signals are atomic propositions extracted from individual Entries. Each must be supported by its source span and retain attribution. Domains provide persistent contextual boundaries, such as a geographic area or organizational relationships, that determine where evidence and queries belong.

Claims preserve interpretations, including their supporting evidence, assumptions and standing. Unlike an ordinary extracted Signal, a derived Claim must say something that no single antecedent entails. Dimensions place Claims on named, ordered axes, such as funding availability or implementation feasibility, allowing comparisons beyond similarity of wording.

The paper’s illustrative example combines evidence about a foundation’s climate funding, a consortium’s presence in Kenyan counties, an existing relationship and a government implementation mechanism. Together, those observations could justify a Claim that partnering with the consortium offers the strongest route to a program. That strategic interpretation is stored with its grounds rather than reconstructed for every subsequent question.

Assumptions receive explicit treatment. If supporting evidence for a precondition exists, it becomes part of the Claim’s justification. If none is found, the assumption remains visible in its context and must accompany downstream use.

Changing support is not the same as disagreement

A dependency network is designed to update Claims when their evidentiary foundations change. A Claim can have alternative justifications: losing one does not remove support if another remains intact.

The architecture distinguishes five standings: proposed, accepted, contested, retracted and superseded. These describe the system’s committed view of support and history, not guarantees of real-world truth. Acceptance can come from a person or a specified automatic rule; workspaces can require human acceptance instead.

Crucially, disagreement and support loss behave differently. A conflict can make two Claims contested without automatically spreading that contest to their dependents. Losing every supported justification retracts a Claim, and that support loss can propagate downstream. Retracted Claims remain available for audit but are excluded from retrieval.

Updates become visible through atomic maintenance batches. This avoids exposing partial propagation, but does not promise immediate freshness: the proposed initial five-minute batching interval is followed by processing time that the architecture does not bound.

What the retrieval experiment found

The reported experiment pooled 272 held-out queries from three workspaces across eight registered retrieval configurations. Its primary measure, nDCG@10, evaluates ranking quality among the first ten results.

The full reranker scored 0.481, compared with 0.453 for a configuration already using the standing penalty. The authors report this Dimension-related increment of 0.028 with p = 0.002. Separate system comparisons, evaluated on 251 queries, report gains of 0.081 over Signals-only retrieval and 0.053 over graph-augmented Signals-only retrieval, both with reported p < 0.001.

Those comparisons do not establish that every architectural component works. In particular, the Dimension comparison holds Claim persistence and the standing penalty constant; it does not test dependency maintenance.

An additional, post hoc comparison removed positional geometry while retaining each Claim’s relevance to the named axes. The full system exceeded that facet-only configuration by 0.017, but the position-specific increment was not statistically significant. The evidence therefore does not establish that geometric position itself explains the retrieval improvement.

The central promise remains untested

The authors explicitly describe the retrieval results as preliminary. Supporting records for bootstrap estimates, sequential testing and original multiple-comparison decisions were unavailable. They also could not reconstruct the reported Dimension-gain interval, which was not centered on the estimated difference.

Pilot annotation exposed another unresolved issue: agreement about whether Claims concerned particular Dimensions was weak or inconsistent. Those observations were descriptive, not held-out validation.

Most importantly, the study did not evaluate truth maintenance under changing evidence, sustained organizational reuse or organizational instinct. Governance over who can read or accept Claims also remains open. Robin Inc. funded the work, and three authors report employment or financial interests in the company, which develops systems based on the architecture.

The contribution is consequently a specified approach to preserving revisable interpretations, accompanied by preliminary retrieval evidence—not proof that an organization can yet rely on a shared, self-maintaining judgment system.