Advanced Settings

Settings for power users who want fine-grained control over Memo's behavior.

Memory / RAG

Controls how the retrieval-augmented generation system fetches context.

Setting	Default	Description
Top-K	5	Number of most relevant memories to retrieve per query. Higher values provide more context at the cost of relevance and latency.
Similarity Threshold	0.7	Cosine similarity cutoff (0.0–1.0). Only memories scoring above this value are returned. Lower values return more results; higher values are stricter.
Min Similarity	0.4	Absolute minimum similarity. Memories below this are never returned regardless of Top-K. Acts as a hard floor.

Model Parameters

Fine-tune the LLM generation behavior.

Setting	Default	Description
Temperature	0.7	Controls randomness in output (0.0–2.0). Lower values produce deterministic, focused responses. Higher values increase creativity and variation.
Repeat Penalty	1.1	Penalizes token repetition (>1.0). Higher values reduce word and phrase repetition at the cost of coherence.
Top-P	0.9	Nucleus sampling threshold. Limits token selection to the most probable cumulative mass.
GPU Layers	-1	Number of model layers to offload to GPU. -1 offloads all layers. Set to 0 for CPU-only inference. Adjust if VRAM is insufficient.

Network & Access

Control how other devices can connect to Memo.

Setting	Default	Description
Remote Access	off	Toggle to allow connections from other devices on the local network. When enabled, API and web UI bind to 0.0.0.0.
API Port	8090	Port for the Memo API server. Must be available and not blocked by a firewall.
Web UI Port	3000	Port for the React web frontend. Change if the default conflicts with other services.
Llama Port	auto	Port for the llama.cpp server. Set manually to pin a specific port, or use auto for dynamic assignment.

Port Configuration Notes

• Ports must be unique across all Memo services and any other running applications.
• If Remote Access is enabled, ensure your firewall allows inbound connections on these ports.
• The llama server uses a random available port by default to avoid conflicts with parallel instances.
• As of v3.3.3, every request against a Remote Access connection (LAN or ngrok) must present the access token shown in Settings — this closes a gap where anyone reachable on the same network could previously read provider API keys or run commands with zero credentials. The mobile app already sends this token automatically.

Developer API Gateway

Sidebar → Developer exposes an Anthropic-compatible local endpoint for tools that only support that protocol (most notably Claude Code, via ANTHROPIC_BASE_URL). Model selection uses a type/model-id format (e.g. local/qwen2.5, openai/gpt-4o).

Setting	Default	Description
Require API Key	off	Shares the Remote Access token; when off, the gateway is open to anything that can reach it locally
Use Memory	off	Routes gateway conversations through RAG memory; never creates a visible chat session even when enabled

Full agentic tool calling (file read/write, run command) works for the openai/custom/local/groq/openrouter/grok/opencode-zen/opencode-go providers. Gemini, Claude, and Ollama's own provider implementations don't support tool calling yet — a tools-bearing request to one of those returns a clear error instead of silently dropping the tools.

Memo Swarm (Beta)

Requires Settings → Beta Features to be enabled first. Sidebar → Swarm lets several PCs pool compute for one oversized GGUF model that wouldn't fit on a single machine.

Setting	Default	Description
RPC Port	50052	Port this machine's rpc-server binds to when acting as a worker ("Join")
Compute Share	manual	Each machine's contribution is set by hand (percentages) — the goal is capacity, not speed

The Host machine keeps the model file and shares a room code; Joiners lend compute without downloading the model. Not available on macOS yet (the helper binary isn't packaged there).