fix(cartesia): keep plain-string search results and skip malformed messages

[Agnik47]

Summary

src/supermemory_cartesia/utils.py and src/supermemory_pipecat/utils.py are copies of the same file. #1434 hardened two spots in the pipecat copy and missed the cartesia one. This brings the two hunks back to parity and adds regression tests.

Bug 1 — plain-string search results are silently dropped

deduplicate_memories → unique_search calls _field(r, "memory", "chunk", "content", default="") on every result. For a plain str that finds no attribute, falls back to "", and the entry never reaches the prompt.

format_memories_to_text in the same file already has an isinstance(item, str) branch for search results, so that branch is currently unreachable — nothing can get past dedup as a string.

Reproduced against the pipecat copy as the reference:

cartesia  search_results -> []
cartesia  rendered       -> ''
pipecat   search_results -> ['User prefers tea']
pipecat   rendered       -> 'Based on previous conversations, I recall:\n\n\n## Relevant Memories\n\n- User prefers tea'

Bug 2 — get_last_user_message raises KeyError

It indexed msg["role"] and msg["content"] positionally, so a tool-call turn or provider event missing either key raised instead of being skipped. Non-string content was also returned verbatim despite the -> str | None annotation.

user with no content   cartesia  -> KeyError: 'content'
user with no content   pipecat   -> 'hello'
entry with no role     cartesia  -> KeyError: 'role'
entry with no role     pipecat   -> 'hello'
non-string content     cartesia  -> [{'type': 'text', 'text': 'hello'}]
non-string content     pipecat   -> None

Changes

• src/supermemory_cartesia/utils.py — both hunks now match the pipecat implementation exactly.
• tests/test_utils.py — new, 6 tests.

Test plan

• 3 of the 6 new tests fail on unmodified main and pass with the fix:

  FAILED tests/test_utils.py::TestGetLastUserMessage::test_skips_entries_without_role_or_content
  FAILED tests/test_utils.py::TestGetLastUserMessage::test_skips_non_string_content
  FAILED tests/test_utils.py::TestDeduplicateMemories::test_keeps_plain_string_search_results
  3 failed, 5 passed
  

• Full suite with the fix: 8 passed.
• ci-python.yml already path-filters packages/cartesia-sdk-python/**, so these run in CI with no workflow change.

No behaviour change for dict or Stainless-model search results — covered by test_still_reads_the_memory_field_of_models_and_dicts.

[Agnik47]

Two things a reviewer would reasonably want to know, both checked:

Do the other SDKs have the same two bugs? No. agent-framework-python and openai-sdk-python already handle both cases — plain-string search results survive dedup, and their get_last_user_message uses .get() with an isinstance guard. cartesia was the only outlier, so this PR closes the class rather than one instance of it.

Is the rest of the file in sync with pipecat now? Yes. The two utils.py copies are behaviourally identical after this change — I differential-tested them across 288 dedup/format input combinations (string / dict / model / empty / unicode / [recent] and [YYYY-MM-DD] prefixed / tag-injection payloads), 8 message shapes, and 4 timestamps: 0 differences.

One thing that audit did turn up, which is not fixed here: both voice SDKs diverge from the TypeScript getMemoryText on a hit shaped {"memory": "", "chunk": "..."} — _field stops at the first non-None value, so it resolves to "" and never falls through to chunk. That one affects cartesia and pipecat equally, so it is out of scope for a cartesia-parity PR. Filed separately as #1693, stacked on this branch.