Providers & models
agent-mem has no API key and no model of its own. It starts the CLI you already use, under your login, with no tools, no MCP servers, no settings and no saved session, in an empty temporary folder.
| Request came from | Model | Command |
|---|---|---|
| Claude Code | Claude Haiku | claude -p --restricted --setting-sources "" --tools "" --strict-mcp-config --no-session-persistence --model haiku |
| Codex | GPT-6-Luna, low effort | codex exec --ephemeral --ignore-user-config --disable hooks --disable memories -s read-only -m gpt-6-luna -c model_reasoning_effort=low |
| Copilot CLI / app | auto, efficiency tier (resolves to GPT-6-Luna) | copilot -s --no-custom-instructions --disable-builtin-mcps --model auto --auto-tier efficiency |
| OpenCode | Big Pickle, free | opencode run --standalone --agent plan --model opencode/big-pickle |
A request never leaves its vendor. A Claude Code request is not sent to OpenAI, a Codex request not to Anthropic. Each uses the cheapest model of its CLI. With only one CLI installed, its requests get facts and the others are saved word for word.
Changing models
Section titled “Changing models”{ "observe": { "claude_model": "sonnet", "codex_model": "gpt-6", "copilot_model": "auto" }}If your plan does not offer a model, or the CLI is too old for the flag, the call is tried once more with the CLI’s
own default. agent-mem doctor and the session panel name the model in use: Facts written by haiku (claude).
Ollama, locally
Section titled “Ollama, locally”{ "observe": { "providers": ["opencode", "claude", "codex", "copilot", "ollama"], "ollama_model": "qwen3.5:4b-mlx" } }Ollama is off by default because local models load the machine. Added, it is the fallback for any request whose own CLI is missing, paused or fails. It is not counted against the daily limit.
Letting any CLI take any request
Section titled “Letting any CLI take any request”{ "observe": { "cross_vendor": true } }Then the first installed CLI in observe.providers takes every request, whatever agent it came from. Off by
default, because it sends one vendor’s conversation to another.
Facts use the subscription or plan you already have, through the CLI’s own login.
| Per batch of up to 5 requests | Calls |
|---|---|
| facts | 1 |
| titles, session summary, lessons | 1 |
| per project and day | 1 review, up to 3 graph batches of 40 memories |
observe.daily_limit (300) caps model calls per day. agent-mem observe --status shows today’s calls per
provider. A usage limit pauses that provider for observe.pause_minutes (60); requests wait rather than being lost.