Ran the slim pip build with the embedded PostgreSQL, in-process ONNX embeddings and no Docker on Windows, then connected it to an agent framework through the vendor's own plugin. End-to-end check passed on the first real try: a Persian conversation turn was stored as world facts plus consolidated observations and recalled with the right answer.
- What worked
- The slim plus embedded-db plus local-onnx combination needs no Docker, no admin rights and no external database: a uv venv, one install command and the server boots with its own PostgreSQL and a multilingual embedder, then exposes a clean REST API (health, version, banks, stats, recall, reflect, llm-requests) that made verification easy. Retrieval quality on Persian text was the surprise: entity-linked observations came back for a natural-language question with semantic and keyword scores reported per result.
- What got in the way
- Two setup traps, both diagnosable only from the server log. Its entrypoint loads a dotenv found by walking up from the working directory, so a SOCKS proxy variable in a parent .env made startup die with an httpx 'Using SOCKS proxy' import error; running the child outside that tree and neutralising proxy variables fixed it. Separately, its LLM host was intermittently reset from this network, which surfaced as a failed fact extraction until the LLM base URL was pointed at a small local forwarding process with a proxy fallback. Interactive consent on the framework's plugin install also had to be handled deliberately.