How to make agentmemory work with Gemini CLI - with Gemini Code Assist Plan

[mansari-lp]

The docs says, agentmemory auto-detects the provider - and we dont need to the LLM Provider Key. I am currently on my company's gemini code assist subscription - but agentmemory is not picking up this info. It is still asking for LLM Provider Key (which i dont have - cause we dont use API KEY in this subscription). How can i fix this ? @rohitg00 - Please help.

The same question for OpenCode as well. How to make this work with OpenCode-GO or OpenCode-Zen subscription ?

[rohitg00]

Hey @mansari-lp — short version: agentmemory does run without any LLM provider key, but in that mode it falls back to a "no-op" provider where memory storage, BM25 search, and recall all work — only the background LLM-driven compression / summarization / lesson-extraction is disabled. The README line is correct for Claude subscriptions (Claude Code IS the LLM), but it's misleading for Gemini Code Assist + OpenCode-Go/Zen users, where the IDE plugin uses OAuth and never exposes a raw key.

Three options today:

1. Run in no-op mode (default). The No LLM provider key found … using no-op provider line is informational, not an error. Your observations are saved, retrieved, and searchable via BM25 + vector recall. You just don't get auto-consolidation / auto-summarization / auto-lessons. For most workflows this is perfectly usable.

2. Get a free Gemini API key from AI Studio — https://aistudio.google.com/apikey. This is separate from your Gemini Code Assist seat: AI Studio has a generous free tier and gives you a raw GEMINI_API_KEY you can drop in ~/.agentmemory/.env:

GEMINI_API_KEY=AIza...

Code Assist (OAuth) and AI Studio (API keys) are different Google product surfaces but both call the same gemini-2.5-flash underneath.

3. Point at a local model. If you run Ollama (ollama serve) or LM Studio, use the OpenAI-compat path:

OPENAI_BASE_URL=http://localhost:11434/v1
OPENAI_API_KEY=ollama        # placeholder — Ollama doesn't validate it
OPENAI_MODEL=llama3.1:8b

The same shape works for any OpenAI-compatible endpoint (vLLM, LM Studio, DeepSeek, SiliconFlow, Azure OpenAI).

For OpenCode-Go / OpenCode-Zen: same situation — those are hosted subscriptions without raw API access, so no provider can hit them directly. Same three options apply (no-op, AI Studio free key, or local model).

Filing a follow-up to tighten the README wording + the stderr warning so this isn't ambiguous for non-Claude users. Thanks for flagging it.