Problem: agentmemory uses a single OPENAI_API_KEY + OPENAI_BASE_URL for both LLM and embedding. When using local Ollama for embeddings (bge-small-zh) and a cloud API (Volcengine/DeepSeek) for LLM, they can’t be configured independently.
Set base URL to Ollama (localhost:11434) → embeddings work, but LLM can’t reach cloud API
Set base URL to cloud (volcengine) → LLM works, but embeddings fail with dimension mismatch
Workarounds tried:
OPENAI_API_KEY_FOR_LLM=false + other provider — other providers don’t support custom base URLs
EMBEDDING_PROVIDER=local — all-MiniLM-L6-v2 has poor Chinese support
Run local LLM (qwen2.5:0.5b/1.5b) — J4125 too slow (40s+ inference)
Proposed solution: Add separate env vars for LLM OpenAI provider, e.g. LLM_OPENAI_API_KEY, LLM_OPENAI_BASE_URL, LLM_MODEL.
Problem: agentmemory uses a single OPENAI_API_KEY + OPENAI_BASE_URL for both LLM and embedding. When using local Ollama for embeddings (bge-small-zh) and a cloud API (Volcengine/DeepSeek) for LLM, they can’t be configured independently.
Set base URL to Ollama (localhost:11434) → embeddings work, but LLM can’t reach cloud API
Set base URL to cloud (volcengine) → LLM works, but embeddings fail with dimension mismatch
Workarounds tried:
OPENAI_API_KEY_FOR_LLM=false + other provider — other providers don’t support custom base URLs
EMBEDDING_PROVIDER=local — all-MiniLM-L6-v2 has poor Chinese support
Run local LLM (qwen2.5:0.5b/1.5b) — J4125 too slow (40s+ inference)
Proposed solution: Add separate env vars for LLM OpenAI provider, e.g. LLM_OPENAI_API_KEY, LLM_OPENAI_BASE_URL, LLM_MODEL.