Explore shared knowledge federation on MemoryNetwork.si, with source lineage, permission-aware retrieval, reversible entity matching and revocation workflows.
A Federation Layer Rather Than an Omniscient Memory
MemoryNetwork.si could connect evidence held in document repositories, issue trackers and application memory stores. Its purpose would be authorized retrieval without flattening every source into an unrestricted repository. A knowledge graph could represent relationships while keeping access dependent on supporting sources and the requesting identity.
.SI is officially Slovenia's country-code top-level domain. Memory and intelligence associations are creative branding, not official meanings or proof of human-like recall. Superintelligence would not be necessary. The proposed service would depend on source quality and reliable permissions rather than the breadth of information collected.
Follow an Evidence-First User Workflow
An engineer investigating a deprecated component could search its name and project. The system would resolve candidate entities and retrieve authorized decisions with dated citations. A graph could show which decision superseded another and distinguish explicit statements from inferred relationships. Restricted incident records should not reveal titles or participants unless policy permits that disclosure.
Incomplete authorized evidence could produce an explicitly limited answer without identifying hidden records. Suggested relationships would enter a maintainer review queue. Accepted merges should remain reversible because similarly named components and teams can cause erroneous links that affect many later queries.
Keep Assertions Separate From Source Facts
Entities, source records, assertions and provenance links could be separate data objects. Assertions would carry source versions, reviewer status and applicable time ranges. Contradictions could coexist rather than being merged prematurely into one supposedly authoritative fact. Confidence in extraction is different from confidence that the underlying statement is true.
The MemoryLogic.si toolkit concept could address policy evaluation across such records, not an existing dependency. The federation product would focus on traceable connections, correction workflows and access boundaries rather than automatically building the largest possible graph.
Authorize Before Material Reaches a Model
Permission checks must apply before candidate evidence enters a generation context, not only when citations render. Connectors could enforce source permissions and recheck them when returning records. Caches need identity and policy context; a shared cache keyed only by query text could leak results between users.
Derived embeddings and summaries may retain information from multiple sources. Lineage tracking could invalidate affected artifacts after permission changes. Immediate revocation across disconnected systems may be impossible, requiring documented freshness bounds and a fail-closed option for sensitive sources. Indexing should not be described as permission preservation without tests.
An MVP and Its Maintenance Procedures
The first version could connect one repository and one issue tracker inside one organization. Read-only retrieval, citations, a limited schema and reviewed entity matches would avoid premature cross-organization complexity. The MindMemory.si personal-context concept illustrates a distinct consent problem; personal preferences should not casually become enterprise evidence.
Operations would include credential rotation, permission-sync monitoring and source-deletion handling. Tests should follow deleted records through indexes, summaries, caches and backups under the stated retention policy. Restoring an old backup must not reinstate material revoked after its creation without reconciliation.
Commercial Viability and Unavoidable Limits
Subscriptions could cover supported connectors, policy administration and private deployment, with transparent indexed-volume limits. Custom connector work can threaten repeatability. A bounded pilot should examine citation correctness, unauthorized disclosure, stale results and the effort of maintaining entity mappings.
Permission drift, fabricated relationships and sensitive inferences remain risks. Passing document-level checks does not resolve every inference concern. The acquisition guide and MemoryNetwork.si listing address naming separately. A launch decision should rest on evidence retrieval without weakening controls, not promises of comprehensive memory or guaranteed compliance.