Preserve identity.
Gate every change.

Freeze model weights and executable code; make every knowledge change reversible.

A verifier and critic use successful and failed traces to propose edits across six modules. A fixed validation gate accepts qualifying patches or restores the prior state. Across 18 domain–model settings, mean holdout score rises from 30.19 to 39.15; 16 improve and two regress.

Evidence & limits. The score measures agreement with logged advertising actions, not revenue or safe live deployment. Each run uses 480 training, 160 validation and 160 holdout cases, not the entire half-million-log source corpus. Main configurations largely run once. Repeated validation selection can overfit, and a passing gate does not guarantee holdout improvement.

Representation and compression are separate design choices.

The system retrieves both stores through semantic search, then consolidates redundant memories offline. On LoCoMo with GPT-4.1-mini, the hybrid scores 90.2 on the paper’s LLM-judge metric, versus 87.1 episodic-only, 83.7 relational-only and 91.3 full-context. Authors report 87% fewer query tokens than full context. One consolidation setting reduces memory size 16.3%, while score falls to 89.2.

Evidence & limits. Query-token savings exclude total curation and consolidation costs. Evaluation uses text conversations, model-dependent extraction and judging; some competing baselines come from earlier work rather than fresh reruns. Full context remains stronger in this setting. These scores are not comparable with the other cards’ metrics.

A new phrase should not automatically create a new memory identity.

Incoming observations attach to durable relationships with source, time and context. Deterministic updates record reinforcement, inactivity and reactivation. Against a registry retaining canonical IDs and semantic fallback but no alias memory, Correct Match Rate improves 7.8–12.1 percentage points across four settings. A blinded audit of 200 incoming observations per setting finds gains of 9–16.5 points.

Evidence & limits. Proprietary investor-call and cash-flow data restrict reproduction. The public earnings-call subset tests reconciliation, not the entire heterogeneous fusion path. Audits cannot measure relationships missed during extraction. ArXiv reports non-archival AI4F workshop acceptance at ACM ICAIF 2026; this was not independently verified.

A provider’s software update can introduce a new AI use without a new procurement.

The guidance covers significant internal and supplier AI use. It describes inventories, risk records, escalation routes and an optional central AI group or committee. Changes to models, data, functionality, access or scale should prompt decisions about further scrutiny and updated controls. Trustees retain accountability for outsourced services.

Evidence & limits. This explains existing duties and expectations for pension schemes; it does not create a separate regulatory regime. Suggested tools are proportionate options, not a universal mandatory committee structure. The publication supplies guidance, not measured implementation results.

Downloadable weights create a concrete alternative to relying on the vendor’s inference API.

The released checkpoint supports self-hosting and customization. Its Mistral Ecosystem Open-Weight License is modified MIT, not unrestricted MIT: a licensee or affiliate operating a model-as-a-service business generally needs separate terms before commercial use. Internal use is exempt; end-user products with capabilities solely embedded in specific features or harnesses are excluded from that business definition.

Evidence & limits. Check the full license for your use. Infrastructure and operational dependencies remain. The vLLM recipe gives substantial accelerator configurations and explicitly does not validate serving at the advertised one-million-token context; its configured cap is 262,144. Vendor coding and security results do not prove safe production operation.