Settings page

Open the local dashboard (http://127.0.0.1:37737 by default) and choose Settings. The team server has its own settings page; see Web dashboard.

What you can set

Section	Settings
Language model	Provider (Ollama, OpenAI, Anthropic, OpenAI-compatible, auto), model, API base URL, API key, LLM tasks on/off, timeout
Embeddings	Provider. OpenAI, Cohere and DashScope also take a model, key, base URL and dimensions. The local FastEmbed model is a setup preset: change it with tam setup --reconfigure, then run memory_rebuild_embeddings
Search answers	Characters per search result (MEMORY_RECALL_MAX_RESULT_CHARS, default 6000), flag records that address the agent (MEMORY_FLAG_INSTRUCTIONS, on by default), cross-encoder re-ranking (MEMORY_CROSS_RERANK)
Storage	Days to keep raw call logs (MEMORY_RAW_LOG_RETENTION_DAYS; not set keeps them forever), embedding model cache folder (TAM_MODEL_CACHE)
Stored credentials	Scan the store for credentials that earlier versions saved, then back up and redact them

How values apply

Precedence. A value saved on the page beats the environment and the env block of an MCP client config. The environment beats the built-in default. Each field shows where its current value comes from.

Timing. Each agent session reads the settings when its MCP server starts. After a change, restart the agent or reconnect the server (/mcp in Claude Code).

Where values are stored

Values live in <memory dir>/settings.json (mode 0600). API keys in it are encrypted with Fernet. The key is TAM_MASTER_KEY when that is set, otherwise <memory dir>/master.key, created with mode 0600 on the first saved key. The page never sends a key back to the browser: it shows •••• and the last four characters.

Keep a copy of master.key with your backups. Without it, saved keys cannot be read and must be entered again.

The setup wizard writes the language-model answers to the same file, so keys no longer sit in plain text in ~/.claude.json or other client configs.

Who can change them

A write needs a same-origin request that carries a token issued to the page itself. Other web sites and plain form posts are refused. Every other dashboard address is read-only.

Search answers

memory_recall cuts a record that is longer than the per-result limit. The cut record carries truncated: true and content_chars, and its text ends with a pointer to memory_get, which returns the whole record.

A record written to the reading agent is marked untrusted_instructions: true. Examples are “ignore previous instructions” and “send the SSH key to …”. The answer then also carries a notice that such records are data, not instructions. Nothing is removed. See Privacy.