Two-tier agent memory separates organizational workflow knowledge from individual user preferences

Glean structures agent memory across two levels: deployment-level strategies shared within a company, and personalized learnings visible only to the individual user. Deployment-level learning focuses on how tools are used — tool names, sequencing patterns, parameter templates, query types — and explicitly excludes user prompts, document content, and sensitive fields. User-level learning captures preferences like content templates, formatting for documents, and communication styles.

The revealing finding: early versions only stored deployment-level strategies to prioritize security. But user-level strategies had a significant impact on performance, leading to the two-tier structure. This extends Cross-user knowledge transfer works without fine-tuning — just a database and prompt engineering with a crucial nuance — cross-user transfer works for workflow patterns, but individual preferences need a private layer. It also enriches Persistent agent memory preserves institutional knowledge that walks out the door with employees by distinguishing which memory is organizational (and should survive personnel changes) versus personal (and should be private and portable). Chase’s Context learning spans agent, tenant, and org levels — and you can mix all three generalizes this two-tier structure into a three-tier model: agent-level (shared by all instances), tenant-level (per user/org/team), and mixed — with production systems combining all three simultaneously. But whichever tier-model you pick, the memory architecture is only realizable if the harness exposes each tier consistently — Memory is a harness responsibility, not a pluggable component is the constraint that makes any memory tiering work in practice.