179 — Tencent ContextPilot: local memory implementation, illustrative demo
Reviewed September6,2026 UTC.

Finding
Official Tencent/ContextPilot implements explicit context and memory tools. Its linked 'live trace' is expressly illustrative, not a captured model run. The inspected memory class stores structured notes, embeddings and relationships in Python dictionaries, with optional external embedding calls. This is relevant Chinese-associated memory research, not evidence of a public scratch-memory swarm.

Primary sources
https://github.com/Tencent/ContextPilot
Complete tree782cbb6611fb610c4cf6fafda6022b7e89cae191,1246entries,nontruncated.
Four pinned source/document files downloaded and Git-blob verified:README.md,infer/README.md,infer/tools/context-shaper_tools.json,train/verl/tools/contextpilot_memory.py.
The official namespace establishes repository provenance. It does not identify an operator for the unexplained paste corpus.

Tools and storage
The schema contains18tools: plan,analyzeText,buildIndex,checkBudget,readChunk,readMultiChunks,searchEngine,memorize,loadMemory,updateMemory,note,readNote,updateNote,deleteContext,truncateContext,summarizeContext,compressContext,finish.
ContextPilotMemoryState holds notes, signatures, embedding vectors and graph edges in process state. It supports addition, reading and append/overwrite/delete updates, with related-memory retrieval. An optional OpenAI-compatible embedding client is configured through environment variables. No actual environment values were read or used. An embedding API call is not publication to a public memory host. The inspected class does not itself implement a publicly writable note site or cross-account swarm communication.
https://github.com/Tencent/ContextPilot/blob/782cbb6611fb610c4cf6fafda6022b7e89cae191/train/verl/tools/contextpilot_memory.py
This class-level scope does not exclude other network or persistence behavior elsewhere in the framework.

Trace check
https://tencent.github.io/ContextPilot/
Direct200 page says the interactive execution trace is an illustrative local trace, with no model endpoint, document upload or API key used in the browser demo. It labels its scenarios ILLUSTRATIVE RUN. Therefore its displayed actions/token counts are not independently observed agent executions. No demo actions or backend submissions performed by this investigation.
README says evaluation writes predictions,trajectories,scores under infer/results/. The complete inspected Git tree contains no infer/results/ paths. Benchmark inputs are present or externally linked, but input data and a trajectory viewer are not actual run records. No full-checkpoint evaluation was run.

Research context and chronology limits
README links an August2026 arXiv identifier2608.28476 and a model collection. The paper/checkpoints were not inspected in this bounded pass, so no claim about training dates, released model quality or past runtime behavior follows. The project describes context-aware branching during training; such alternative rollouts are not automatically a communicating deployment swarm.

Discovery
Two targeted Tencent/Youtu/AdaSkill searches surfaced this official project along with other research candidates. Followed this lead because structured memory and explicit context edits are closer to the original scratch-memory question than generic role-team marketing.

Next
Seek published raw training/evaluation rollouts, not illustrative demos. Tencent Youtu-Agent's advertised future trajectory release and CognitiveKernel-Pro's session serialization are separate candidates for follow-up. Preserve the distinction between model context edits, local persistent notes, external embedding calls and public writable storage.

Evidence
179-private:tree.json/hash manifest; four pinned blobs and sources.json;demo.html and demo-capture.json. Six direct public GETs,two searches. No package installation, source execution, evaluation, credential access or external write. No escaped Chinese swarm confirmed.