STM Sub-Pipeline

Disclaimer

This specific page uses temporary drafts written and maintained by Generative AI. While verified to be accurate, please cross-verify with source code.

The short-term memory (STM) sub-pipeline maintains working memory across recent conversation turns. It stores recent dialogue in an in-process cache and records model-generated summaries and structured categories in the database.

The sub-pipeline provides two write paths:

Stage	Trigger	Written state	Persistence
01: Passive Capture	Post-turn, generation finished	Recent dialogue turns up to MAX_MESSAGES_PER_CHANNEL	In-process cache only; ensures scope identity row in database
02: Summary Upgrade	Tool loop, update_short_term_memory call	Distilled summary text or structured category key-value pairs	In-process cache and durable short_term_memories table

Key design facts

• Hybrid persistence: crude messages exist only in the in-process cache, while summaries, categories, and cadence state are stored in the short_term_memories table. Following a process restart, durable entries are hydrated with an empty crude message list, and message history repopulates as turns occur.
• Dual-key scoping: writes index into both user-scoped (shortterm:user:userId:channelId[:personaId]) and server-scoped (shortterm:server:serverId:channelId[:personaId]) keys. Direct message sessions create only the user-scoped entry.
• Summary priority: same-channel context favors categories or summary text. Other-channel context can include raw messages alongside distilled content according to render mode. Passive capture preserves existing summaries.
• Single upgrade per turn: streamingContext.disableShortTermMemoryUpdate is set to true after the first successful upgrade, preventing the model from invoking update_short_term_memory multiple times in one tool loop.
• Deletion ownership: channel resets through /refresh call clearShortTermMemoryForChannel(), and user memory purges call clearShortTermMemoryForUser(). Both helpers evict the in-memory cache and delete rows from the short_term_memories table in the background.

Cross-references

• Intent detection gate: Memory Pipeline explains how explicit memory intent suppresses the STM upgrade tool.
• Context assembly consumer: Short-Term Memory Context formats cached messages and summaries for prompt injection.
• Tool-loop caller: Execute Tool Call dispatches tool invocations that trigger summary upgrades.
• Post-turn effects caller: Run Generation Turn: Post-Turn Effects triggers passive turn capture.