PATCH re-joins topics/entities arrays with pipe separator, corrupting TAG fields

[chrisbaker2000]

Description

When updating a memory via PATCH /v1/long_term_memory/memories/{memory_id}, the topics and entities array fields get re-joined with pipe (|) separators into a single string element, corrupting the TAG field data in Redis Search.

Steps to Reproduce

1. Create a memory with separate topics:

curl -X POST ".../v1/long_term_memory" \
  -H "Content-Type: application/json" \
  -d '{"text": "test", "topics": ["cooking", "italian"], "entities": ["pasta"]}'

2. Update the memory text via PATCH:

curl -X PATCH ".../v1/long_term_memory/memories/<id>?namespace=default" \
  -H "Content-Type: application/json" \
  -d '{"text": "updated test"}'

3. Fetch the memory:

curl ".../v1/long_term_memory/memories/<id>?namespace=default"

Expected: topics: ["cooking", "italian"]
Actual: topics: ["cooking|italian"]

Root Cause

The PATCH handler appears to read the existing TAG field values (which use | as the internal Redis Search separator), then re-store them as a single pipe-joined string rather than splitting them back into an array.

Impact

• Progressive corruption: Every PATCH compounds the issue — after 2 patches: ["cooking|italian|cooking|italian"]
• Filter breakage: Topic/entity filter queries stop matching because the TAG value is now a single combined string
• Affects all PATCH operations: Even patches that don't touch topics/entities corrupt them as a side effect
• Scale: In our deployment (~3,500 memories), 236+ memories have pipe-corrupted topics/entities from routine curator PATCH operations

Environment

• Server version: 0.13.2
• Redis Search index with TAG fields for topics and entities

Suggested Fix

When reading existing memory data for a PATCH merge, split TAG field values on | before re-storing:

# When reading existing topics/entities from Redis
existing_topics = existing.topics.split("|") if isinstance(existing.topics, str) else existing.topics

Or ensure the PATCH code path uses the same array-aware serialization as the POST path.

Workaround

We run a weekly curator that detects and repairs pipe-corrupted entries by splitting on | and re-storing as proper arrays. However this is a lossy process and doesn't prevent re-corruption on the next PATCH.