fix(memory): flatten multimodal content before provider sync

[kshitijk4poor]

Salvage of #44517 by @erosika — original fix cherry-picked verbatim (authorship preserved), plus one follow-up commit that dedupes the helper. Based on current main.

Problem (commit 1, @erosika)

Turns with an attached image log this and silently never reach the memory backend:

WARNING agent.memory_manager: Memory provider 'honcho' sync_turn failed: expected string or bytes-like object, got 'list'

Multimodal user messages carry content as a list of typed parts ({type: "text"|"image_url", ...}). _sync_external_memory_for_turn passed that list straight into MemoryManager.sync_all, and providers feed it to regexes — Honcho's sync_turn calls sanitize_context in its synchronous prologue, where re.sub raises the TypeError before the sync thread spawns (caught by sync_all, surfaced as the warning above). mem0, supermemory, and the other providers expect a string the same way.

The fix flattens content to plain text at the boundary in _sync_external_memory_for_turn, before sync_all/queue_prefetch_all — one fix covers every provider instead of patching each plugin:

• strings pass through, None → ""
• text parts joined, images become a [N image(s)] marker so the attachment isn't erased from recall
• turns with no recoverable text are skipped rather than synced empty

Both original_user_message and final_response are guarded — both are typed Any.

Follow-up (commit 2): reuse the existing helper instead of forking it

The original commit added flatten_message_content to agent/memory_manager.py, but it was a near-exact duplicate of agent/codex_responses_adapter.py:_summarize_user_message_for_log — same None/str/list dispatch, same {text,input_text,output_text} / {image_url,input_image} part sets, the identical [N image(s)] marker, and the same scalar fallback. The only difference was the join separator (newline for memory vs the space-join the existing helper already uses for log/trajectory previews), and that helper is already imported into agent/turn_finalizer.py — the same file whose call site the memory fix touches.

This commit parameterizes the existing helper with sep=" " (default preserves every current logging/trajectory caller byte-for-byte) and calls it with sep="\n" at the memory boundary; the forked flatten_message_content is dropped. Single source of truth for multimodal-content flattening, no behavior change for the fix or for existing callers.

Tests

• tests/agent/test_memory_provider.py — unit coverage for the consolidated helper (newline-join for memory, plus a case locking the default space-join), including a regression case asserting the flattened output survives sanitize_context
• tests/run_agent/test_memory_sync_interrupted.py — boundary coverage: multimodal user message, multimodal response, and no-recoverable-text skip

All touched suites pass locally (110 tests), ruff clean. Verified end-to-end on the branch: a multimodal list flattens to [1 image] <text> and survives sanitize_context; a raw list still crashes it (the bug); logging callers keep the space-join.

Salvage of #44517.