STM Sub-Pipeline

Manages the short-term memory cache — an in-process Map that stores the last 10 conversation turns per channel/persona pair. Two write paths exist:

Stage	File	Trigger	What it writes
01-passive-capture.md	storeShortTermMemory	Post-turn, always	Crude conversation (last 10 messages + new persona response)
02-summary-upgrade.md	updateShortTermMemorySummary	Mid-turn, LLM tool call	LLM-authored summary that replaces the crude conversation

Key design facts

• Cache-only — STM is never written to the database. A process restart clears all STM entries. Cross-restart persistence is not a design goal.
• Dual-key scoping — every write creates (or updates) two cache entries: one user-scoped (shortterm:user:userId:channelId[:personaId]) and one server-scoped (shortterm:server:serverId:channelId[:personaId]). DM sessions get only the user-scoped key.
• Summary takes priority — the context-build stage renders the summary field when present, ignoring the messages array. The passive capture (stage 01) preserves any existing summary when overwriting the messages array.
• One upgrade per turn — streamingContext.disableShortTermMemoryUpdate is set to true after the first successful update_short_term_memory tool call, preventing the LLM from calling it again within the same tool-loop iteration.

Cross-references

• Intent gate that may suppress stage 02: → memory pipeline README
• Read side (both stages): → context-build STM stage
• Tool-loop that triggers stage 02: → tool-loop Stage 02 executeToolCall
• Post-turn effects that trigger stage 01: → chat per-turn Stage 04