feat(pipecat): add Pipecat voice AI pipeline memory integration

[benfrank241]

Summary

Adds hindsight-pipecat — a FrameProcessor for Pipecat voice AI pipelines that adds persistent memory via Hindsight.

• HindsightMemoryService(FrameProcessor) — slots between the user context aggregator and LLM service
• On each OpenAILLMContextFrame/LLMContextFrame: retains new conversation turns (fire-and-forget), recalls relevant memories (awaited), injects as <hindsight_memories> system message
• Pure asyncio — uses arecall/aretain async client methods, no thread pool
• Handles both OpenAI and base LLM context frame types
• Multimodal content extraction (text parts from list content)
• Modeled after Pipecat's existing Mem0 memory service

Tests

17 tests, all passing:

• Config resolution and error handling
• Recall injection, empty result handling, error swallowing
• Retain turn-pair extraction, idempotency, error swallowing
• Memory message replacement (no duplication across turns)
• Disabled features (no client calls)
• Multimodal user message text extraction

Checklist

• hindsight-pipecat package with HindsightMemoryService
• 17 unit tests (all passing)
• CI job in .github/workflows/test.yml
• Entry in scripts/release-integration.sh VALID_INTEGRATIONS

🤖 Generated with Claude Code