Chase identifies three granularities at which context learning happens: agent-level (the agent updates its own persistent configuration — OpenClaw’s SOUL.md that evolves over time), tenant-level (each user, org, or team gets their own context that updates independently — Hex’s Context Studio, Decagon’s Duet, Sierra’s Explorer), and mixed (an agent can have agent-level + user-level + org-level context updates simultaneously).
This directly extends the deployment/user split in Glean’s trace learning system (below). Chase’s framework adds the agent level as a third tier — the agent itself has persistent configuration that evolves independently of any tenant. Your own setup is a living example: CLAUDE.md skills and memory operate at the “tenant” level (you, Ayush), while Claude Code’s built-in behaviors are the “agent” level, and Anthropic’s model improvements are the “model” level. The implication for Cross-user knowledge transfer works without fine-tuning — just a database and prompt engineering is that transfer happens at the tenant level — one user’s context learnings can benefit others at the org level without touching the model or harness.
Glean’s version (@tonygentilcore — “Trace Learning for Self-Improving Agents”, sources/tonygentilcore-trace-learning-self-improving-agents.md) splits memory into deployment-level strategies shared within a company and personalized learnings visible only to the individual user. Deployment-level learning covers 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, document formatting and communication styles. The revealing finding: early versions stored only deployment-level strategies to prioritize security, but user-level strategies had a significant impact on performance, which led to the two-tier structure. Which memory is organizational (and should survive personnel changes) and which is personal (and should stay private) is a design decision, and Memory is a harness responsibility, not a pluggable component is the constraint that makes any tiering work in practice.