Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.
Python agentaianthropicclaude-codecompressioncontext-engineeringcontext-windowcursorfastapilangchainllmmcpopenaiprompt-engineeringproxypythonragtoken-optimizationtokenstypescript
fix(memory): keep multimodal user content intact when injecting memories (#4648)
HetCreep07d5add5ba31
diff --git a/headroom/memory/wrapper.py b/headroom/memory/wrapper.pyindex 0f55d35df..7053a58d1 100644--- a/headroom/memory/wrapper.py+++ b/headroom/memory/wrapper.py@@ -27,6 +27,19 @@ from headroom.memory.models import Memorylogger = logging.getLogger(__name__)+def _text_of(content: Any) -> str:+ """Return the text of a message's content: a string, or the text parts of a list."""+ if isinstance(content, list):+ return "\n".join(+ part["text"]+ for part in content+ if isinstance(part, dict)+ and part.get("type") == "text"+ and isinstance(part.get("text"), str)+ )+ return str(content)++class MemoryWrapper:"""Wraps an LLM client to add automatic memory with zero extra latency.@@ -140,12 +153,18 @@ class MemoryWrapper:if not user_content:return messages+ # Multimodal content is a list of parts; search on its text parts only,+ # never on the list's repr.+ query = _text_of(user_content)+ if not query:+ return messages+# Search for relevant memories (async -> sync)loop = asyncio.new_event_loop()try:memories = loop.run_until_complete(self._memory.search(- query=str(user_content),+ query=query,user_id=self._user_id,session_id=self._session_id,top_k=self._top_k,@@ -169,7 +188,12 @@ class MemoryWrapper:for msg in new_messages:if msg.get("role") == "user":original = msg.get("content", "")- msg["content"] = f"{context_block}\n\n{original}"+ if isinstance(original, list):+ # Keep the parts (images, files) as they are; add the+ # context as a leading text part.+ msg["content"] = [{"type": "text", "text": context_block}, *original]+ else:+ msg["content"] = f"{context_block}\n\n{original}"breakreturn new_messagesdiff --git a/tests/test_memory_wrapper.py b/tests/test_memory_wrapper.pyindex 5907766a3..02b24922d 100644--- a/tests/test_memory_wrapper.py+++ b/tests/test_memory_wrapper.py@@ -121,6 +121,46 @@ def test_inject_memories_handles_empty_and_inserts_context() -> None:}+def test_inject_memories_keeps_multimodal_content_parts() -> None:+ """List (multimodal) user content keeps its parts; the context is a new text part.++ See https://github.com/headroomlabs-ai/headroom/issues/4362 (item 10).+ """+ client, _response = make_client()+ fake_memory = FakeMemory()+ wrapper = MemoryWrapper(client, user_id="alice", _memory=fake_memory)+ fake_memory.search_results = [+ SimpleNamespace(memory=SimpleNamespace(content="Prefers Python")),+ ]+ image = {"type": "image_url", "image_url": {"url": "data:image/png;base64,AAAA"}}+ parts = [{"type": "text", "text": "What is in this chart?"}, image]+ original = [{"role": "user", "content": parts}]++ injected = wrapper._inject_memories(original)++ assert original[0]["content"] == parts+ assert injected[0]["content"] == [+ {"type": "text", "text": "<context>\n- Prefers Python\n</context>"},+ {"type": "text", "text": "What is in this chart?"},+ image,+ ]+ assert fake_memory.last_search["query"] == "What is in this chart?"+++def test_inject_memories_skips_search_for_content_without_text() -> None:+ client, _response = make_client()+ fake_memory = FakeMemory()+ wrapper = MemoryWrapper(client, user_id="alice", _memory=fake_memory)+ fake_memory.search_results = [+ SimpleNamespace(memory=SimpleNamespace(content="Prefers Python")),+ ]+ image = {"type": "image_url", "image_url": {"url": "data:image/png;base64,AAAA"}}+ messages = [{"role": "user", "content": [image]}]++ assert wrapper._inject_memories(messages) == messages+ assert not hasattr(fake_memory, "last_search")++def test_store_memories_persists_only_nonempty_content() -> None:client, _response = make_client()fake_memory = FakeMemory()