Thanks — I hear you: you want a way to automatically compress/summarize old function-call history (keep recent pair intact, replace older FunctionCall with one-line summaries and drop heavy FunctionCallOutput payloads) to reduce tokens, latency, and cost. Short answer: LiveKit does not have a built‑in automatic feature that compresses or summarizes conversation history for you. What you can do instead • Implement client/server-side pruning and summarization before you build the prompt: ◦ Keep all user/assistant messages intact (as you said). ◦ Replace older

FunctionCall

events with a one-line summary (who/what/intent/time). ◦ Strip out old

FunctionCallOutput

payloads entirely, or store them externally and reference them by ID. ◦ Always keep the current question’s FunctionCall/Output pair intact. • Use a ConversationManager helper pattern (or similar middleware) to manage what gets sent to the LLM: ◦ Maintain the full history in

session.history

(or your own store). ◦ When preparing

chat_ctx

for the LLM, build a reduced context that includes: ▪︎ Recent messages (last N turns). ▪︎ System message summarizing earlier turns (one-line or short bullet list). ▪︎ Summaries for old FunctionCalls in place of full events. ◦ Optionally include pointers/IDs to the stripped outputs in case you need to rehydrate them later. Practical tips to implement 1. Decide retention rules: ◦ e.g., keep last 3 Q/A turns fully, summarize anything older than that. 1. Create concise summary format: ◦ For FunctionCall: "Q5: validated_user_email -> validated (params: {email_masked})" ◦ For FunctionOutput removed: "Output for Q3 stored at output_id:12345" 1. Store full outputs in external storage (DB or object store) and reference by ID to avoid token costs. 2. Run periodic summarization: ◦ After every X questions, append a short summary system message (1–2 lines) representing the combined earlier context. 1. Measure and iterate: ◦ Track tokens, TTFT, and LLM cost to tune retention/summarization thresholds. If helpful, I can: • Sketch a small pseudocode flow for a ConversationManager that filters/summarizes history before calling the LLM. • Suggest concrete summary templates for your FunctionCall and FunctionCallOutput items. • Recommend thresholds (how many turns to keep vs summarize) based on your 37-question workflow. Which would you prefer next: a pseudocode flow you can drop into your agent, or a few concrete summary templates to use right away? Sources: Building multi-agent architectures with LiveKit agents | LiveKit