Layered long-term memory ported from TencentDB-Agent-Memory's MemoryCore pipeline: conversations are auto-distilled into L1 facts, L2 scene summaries and L3 persona, then recalled and injected before each step. Hybrid retrieval (SQLite FTS5 BM25 + sqlite-vec vectors with optional local embedding models), per-session capture modes (auto/chat/work/off), recall dedupe and recency decay, settings UI with memory browser, and rebuild from the L0 raw-event source.
dsh plugin --profile web add dsh-layered-memory