Problem

AMS stores long-term memory topics, entities, and extracted_from as comma-delimited TAG strings in
Redis.

That means if a single tag value itself contains a comma, it can be split incorrectly when read back.

Example:

["cooking, italian style", "recipes"]

can be decoded as:

["cooking", "italian style", "recipes"]

instead of the original 2 values.

Decision

We should reject commas in submitted tag values rather than add custom escaping/encoding logic.

Why this is the right choice

• simple to implement
• easy to explain
• avoids silent corruption
• matches the TAG separator defined in the Redis index
• avoids introducing a custom encoding format into AMS

What we should do

• validate topics, entities, and extracted_from on write/update paths
• reject values containing commas
• return a clear validation error
• document the restriction

What we should not do right now

We should not add a custom escaping or sentinel encoding scheme unless we decide we truly need arbitrary tag
strings.

That would require coordinated changes across:

• storage encoding
• decoding
• filter/query generation
• migrations/backfills
• tests

Risk if we do nothing

Comma-containing tag values can be silently corrupted and later read back incorrectly.