The pipecat, cartesia, and agent-framework packages all pass
search results from the Supermemory SDK (pydantic models, attribute
access, snake_case fields) into dedup/format helpers written for plain
dicts. The consequences differed by package but all killed the
feature's primary path:
- pipecat: deduplicate_memories called r.get() on a Result model ->
AttributeError -> the outer handler logs "Error processing frame"
and forwards the frame unchanged, so memories are never injected for
any user whose profile lookup returns search results
- cartesia: identical crash, swallowed by _enrich_event_with_memories'
generic except -> memory_context silently comes back empty and the
previous memory block is stripped from the system prompt
- agent-framework: extract_memory_text only handled dict/str, so model
objects fell through to `return None` and every search-result memory
was silently dropped from the injected context
(The OpenAI package shares the same helper shape but works because it
fetches profile data over raw aiohttp and receives dicts — that
asymmetry is what hid this.)
pipecat/cartesia gain an extract_search_result_fields helper that reads
memory/updatedAt off dicts (camelCase keys) or SDK models (snake_case
attributes, datetime-tolerant), used by both deduplicate_memories and
format_memories_to_text; the formatter also no longer prints raw object
reprs for non-dict items. agent-framework's extract_memory_text gains a
model-object branch.
Verified against the real SDK: built a
supermemory.types.search_memories_response.Result from an API-shaped
payload and ran it through dedup -> extract -> format (crashed with
AttributeError before, renders "- [1 Jan] User prefers async" after).
Test suites: pipecat 7 passed, cartesia 7 passed, agent-framework 56
passed (agent-framework-core pinned to 1.0.0rc3 — newer releases have
dropped BaseContextProvider, which breaks the package import
independently of this change).
#### RE-RAISING Pipecat live speech PR
### Added native speech-to-speech model support
### Summary:
- Speech-to-speech support - Auto-detect audio frames and inject memories to system prompt for native audio models (Gemini Live, etc.)
- Fix memory bloating - Replace memories each turn using XML tags instead of accumulating
- Add temporal context - Show recency on search results ([2d ago], [15 Jan])
- New inject_mode param - auto (default), system, or user
### Docs update
- Update the docs for native speech-2-speech models
