v0.21.0

Layered agent memory — every agent grows its own knowledge graph, distills it into a bounded single-read layer, and managers read only the distilled layer of their reports, recursively up an org tree. No reader ever needs the whole archive; no memory outgrows a context window. Entirely opt-in and additive.

Highlights: tesserae agents (identity + org registry), tesserae distill (per-agent L1 artifacts with provenance-preserving forgetting), agent= scoping on every MCP read tool (worker/manager/org views) plus agent_view_explain and audit-logged drill_down, recursive L2' manager rollups, and the growth loop (per-agent harness + purpose.md, guidance streams, semantic bridges, topic maps, subagent promotion). Determinism held throughout: corpus-clock time, lineage-key identity, byte-idempotent artifacts.

Full notes: https://github.com/ca1773130n/Tesserae/blob/main/docs/release-notes/v0.21.0.md · Guide: https://github.com/ca1773130n/Tesserae/blob/main/docs/agent-memory.md