Liu ZhuoQi

Almost every agent project now claims to provide “long-term memory.” For one project, that means embedding chat history. For another, it means maintaining a user profile. A third lets the model edit Markdown files. A fourth builds a bitemporal knowledge graph. All four use the word memory, but they are not the same system and should not be placed on one undifferentiated leaderboard. To decide whether a system genuinely remembers, I would rather ask three questions:

Aliyun’s CAP has a piece on picking an inference engine that narrows the field to four: Ollama, vLLM, SGLang, and Hugging Face Pipeline. In 2024, that framing was fine. By 2026, it’s missing half the map. NVIDIA’s TensorRT-LLM has completed its “PyTorch-ification,” SGLang became famous as the first open-source project to reproduce DeepSeek’s large-scale deployment, Hugging Face slapped a “maintenance mode” banner on TGI and told you to switch to vLLM — and the real throughline of the entire 2025 inference landscape can be summed up in one word: disaggregate.

OpenClaw’s daily-ai-news cron job kept timing out. The root cause: a missing absolute path in the SKILL.md caused the Agent to spend 15 exec calls searching for a tool every run. Messages 165→54, exec calls 44→7 — one file path beat any algorithm optimization.

OpenClaw’s vector retrieval silently failed — but BM25 text search kept the memory system running for two weeks unnoticed. Should you even bother fixing it? Here’s how I used NVIDIA’s free embedding API to complete the picture at zero cost.

The important change is not “two more tool features.” It is the movement of multi-step orchestration into a code-execution environment, with only a compact result returning to model context. Background: The Cost Problem in Agent Tool Calling # In traditional agent tool-calling, every tool invocation requires a full cycle of “model inference → tool execution → result return → model re-inference.” This seemingly natural loop breaks down at scale in three ways: