Memory Connector Rule: Test Export Before You Standardize
Should an agency standardize its agent stack on a memory or knowledge connector now, and what must be proven before that layer holds client institutional knowledge? Run a full export-and-reimport test on real client memory before standardizing any connector, and reject the layer if the graph, notes, and provenance do not survive the round trip intact.
By InnovaAI ResearchPublished
“Should an agency standardize its agent stack on a memory or knowledge connector now, and what must be proven before that layer holds client institutional knowledge?”
Run a full export-and-reimport test on real client memory before standardizing any connector, and reject the layer if the graph, notes, and provenance do not survive the round trip intact.
Operators judge a connector on recall speed and demo quality, then discover at offboarding that the entity graph, freshness metadata, and proof IDs export as flat text or not at all, forcing the delivery team to rebuild client context by hand.
The category description frames the strategic upside as compounding knowledge and the core risk as lock-in, so portability is the deciding test rather than recall quality. Architectures differ sharply: Cogni retrieves through an entity graph with spreading activation and no LLM in the retrieval path, while Knownbase stores decisions and constraints as searchable project notes, and a layer that cannot move between those models strands the agency's institutional record. Vendor terms are also unstable, with 105 of 175 AI tools changing monetization-relevant terms in a 15-week window, so a memory store that cannot be extracted is a liability the moment a tool is repriced or retired.
- •Two or more agents (coding, research, content) touch the same client account and re-prompt for the same brand rules or architecture decisions each session
- •The agency is consolidating memory across tools after a vendor term change or a tool swap, as happened to 105 of 175 AI tools between June and October 2026
- •A retainer depends on continuity across long-running automation, where a forgotten constraint means rework billed against fixed fees
- •Client work involves regulated or auditable claims, so every recalled fact needs a provenance trail
- •The team is choosing between graph-based recall and structured project archives as its primary memory model