Feature Catalog
Memo ships with a comprehensive set of local-first AI capabilities. Every feature runs on-device by default with opt-in cloud extensions.
Intelligence & Memory
| Feature |
Status |
Description |
| RAG Semantic Memory |
Stable |
768-dim vector embeddings, chunked document ingestion, cross-mode recall |
| Contextual Recall |
Stable |
Automatic context injection from past conversations and ingested documents |
| Infinite Context |
Stable |
Chunked document storage with lazy loading; no token limit on stored knowledge |
| Cross-Mode Memory |
Stable |
External chat memory + local embedding server — single vector store for all sources |
| Incognito Mode |
Stable |
Zero-persistence sessions — no writes to disk, no vector indexing, no history |
Model Management
| Feature |
Status |
Description |
| The Factory / Discover |
Stable |
Built-in model store with HuggingFace search, browse, and download; OR-combined multi-select filters (Tools/Vision/Code/Embedding/Size) |
| Background Downloader |
Stable |
Multiple concurrent downloads (no longer one-at-a-time), combined progress in the engine status bar |
| System Diagnostics |
Stable |
Auto-detects NVIDIA VRAM, AMD VRAM, system RAM; first-run setup recommends a matching chat + memory model pair |
| Hardware Compatibility Badges |
Stable |
At-a-glance indicators: GPU-accelerated, CPU-only, RAM requirements |
| Quantization Labels |
Stable |
Plain-language model quality descriptions (Q4_K_M → "Great quality, fast") |
| Two-Panel Layout |
Stable |
Downloaded models + HuggingFace catalogue side by side |
| Real Capability Detection |
Stable |
Tool-calling/code badges derived from the model's actual chat template and tags, not a hardcoded list |
| Context Size Guard |
Stable |
Reads the model's real max context from the file; the slider can't be pushed past it |
External Providers
| Provider |
Auth Type |
Notes |
| OpenAI |
API Key |
GPT-4o, GPT-4, O-series |
| Anthropic Claude |
API Key |
Claude 3.5 Sonnet, Claude 3 Opus |
| Google Gemini |
API Key |
Gemini 1.5 Pro, Gemini 2.0 Flash |
| xAI Grok |
API Key |
Grok-2, Grok-3 |
| Groq |
API Key |
Ultra-fast inference (LPU hardware) |
| OpenRouter |
API Key |
Unified access to 200+ models |
| Ollama |
Local (no auth) |
Any GGUF model, local inference |
| OpenCode Zen |
API Key |
Pay-as-you-go gateway, some models free; pick from a live model list |
| OpenCode Go |
API Key |
Subscription-based gateway; pick from a live model list |
| Kilo Code |
API Key |
AI Gateway, 368+ models, some free; pick from a live model list, no key needed to browse (new in v3.9.0) |
| Custom (OpenAI-compatible) |
Base URL |
Any OpenAI-compatible endpoint — LM Studio, vLLM, your own proxy |
| Custom (Anthropic-compatible) |
Base URL |
Any Anthropic Messages API-shaped endpoint |
| Claude Code / Codex (CLI) |
None |
Shells out to a locally installed claude/codex CLI as a subprocess, per-chat, real background job (Beta) |
| gemini-sub |
Google OAuth |
Beta — sign in with a personal Google account, reach Gemini via Code Assist on your own AI Pro/Ultra quota, no separate API key |
All provider API keys are stored locally in config/config.yaml. Keys never pass through any intermediary server.
Agent Mode
| Feature |
Status |
Description |
| Built-in Tools |
Stable |
27 tools with JSON schemas: file read/edit/write, run_command, web search, calendar, routines, WhatsApp, provider config |
| Permission System |
Stable |
Danger-level based (safe/medium/dangerous) — safe tools always auto-run; medium/dangerous ones prompt with 6 possible responses (allow once/session/forever, deny once/forever, or a 60s auto-deny timeout) |
| Security Sandbox |
Stable |
Path validation, 43-pattern command blacklist, 30 req/min rate limit |
| Pipeline |
Stable |
Iterative tool-calling loop (max 40 iterations) with an audit trail persisted to disk |
| Web Search |
Stable |
Scoped, single-tool version of the same agent pipeline for plain chat — see Agent Mode |
| Code Mode: Plan / Auto / Build |
Stable (v4.5.0) |
Three sub-modes cycled with Ctrl+Tab, each with its own editable system prompt. Plan investigates and writes a saved plan without touching a file; Auto confirms edits as before; Build runs edits and run_command without waiting. Auto-permission chains a finished plan straight into Build, same reply |
| Self-Driving Task Loop |
Stable (v4.4.0) |
Hand Memo a Task.md checklist and walk away — planner/executor mode with plan approval, up to 3 parallel sub-agents (coder + analyzer/reviewer/test-runner), escalating retry, chat + push notification on every terminal state |
| Provider Requirement |
— |
An external provider is the reliable path; a running local llama.cpp model is also wrapped into the same tool-calling request, but real support varies a lot by model |
Orchestra Mode
| Feature |
Status |
Description |
| Chief + 8 Expert Roles |
Stable |
Specialist agents (researcher, coder, reviewer, writer, analyst, architect, debugger, planner) |
| Three-Phase Workflow |
Stable |
Plan → Execute → Synthesize with dependency graph |
| Parallel Execution |
Stable |
Independent tasks run concurrently via goroutines |
| Progress Streaming |
Stable |
Real-time SSE updates during orchestration |
| Slash Command |
Stable |
/orchestra to invoke from any chat |
| Orchestra + Agent Mode Together |
Stable (v3.5.5) |
Each delegated task now runs through the same real, sandboxed, permission-gated agent pipeline a direct agent chat uses — real tool calls, not simulated ones |
Connectivity & Sync
| Feature |
Status |
Description |
| WhatsApp Bridge |
Stable |
whatsmeow-based, QR code pairing, on-device credentials; self-chat assistant + chat-driven Routines as of v3.9.0 |
| Telegram Assistant |
Stable (v3.9.0) |
Connect a bot token, first-message owner lock, same self-chat assistant and Routines-from-chat as WhatsApp |
| Cloud Backup (Google Drive) |
Stable |
AES-256-GCM encrypted, opt-in, per-session or full backup |
| Remote Access |
Stable |
LAN, ngrok, Tailscale — connect mobile app to desktop backend; every request requires a credential valid for the configured auth mode |
| Self-Hosting |
Stable (v3.5.5) |
Server-only installers, Docker/CasaOS, 4-mode auth, multi-account, full memo CLI toolkit over SSH — see Self-Hosting |
| Per-Account Permissions |
Stable (v3.9.0) |
7 independent switches per self-hosted account (Models, Memory, Agent, Calendar, WhatsApp, Telegram, Routines), enforced server-side |
| Mobile Companion |
Beta |
Flutter Android/iOS app, connects to desktop server over network, fully localized (TR/EN) |
| Developer API Gateway |
Stable |
Anthropic- and OpenAI-compatible local endpoints (Sidebar → Developer) for tools like Claude Code, via ANTHROPIC_BASE_URL, or any OpenAI-SDK tool. Both endpoints now enforce the API key for non-loopback callers (v4.5.0 security fix) — see full walkthrough |
| Memo Swarm |
Beta |
Pool several PCs' compute for one oversized local model (Sidebar → Swarm, requires Beta Features); host/join with a room code; not yet on macOS |
Automation & Self-Awareness
| Feature |
Status |
Description |
| Routines |
Stable |
Scheduled automations in plain language — simple prompt or full agent run, device-local timezone; creatable, listable, and cancellable straight from WhatsApp or Telegram as of v3.9.0 |
| Proactive Learning & Ambient Nudges |
Stable |
On by default (subtle); notices habits/patterns and surfaces them via a suggestion banner or woven into a reply |
Self-Insight (/insight) |
Stable |
Looks back over mood history and memory to describe a real pattern, on demand or via a weekly Routine |
| Usage Stats |
Stable |
Settings → Stats: requests, tokens, tokens/sec, most-used model, 30-day chart, per-model breakdown |
| Minimal Mode |
Stable |
Strips personality/mood/web-search instructions for lowest-overhead local use; independently re-enable-able sub-toggles |
UI/UX
| Feature |
Status |
Description |
| Streaming Token-by-Token |
Stable |
SSE-based, Markdown rendering with syntax highlighting |
| Slash-Command Palette |
Stable |
/agent, /orchestra, /search, /model, /incognito |
| File & Image Drop |
Stable |
Drag-and-drop attachments into chat |
| Dark/Light Theme |
Stable |
System-following with manual override |
| Responsive Layout |
Stable |
Desktop + mobile adaptive UI |
| Desktop Mascot |
Stable (v4.5.0) |
A small always-on-top character, its own window sharing the same running app, reflecting Memo's live activity across chat, WhatsApp, Telegram, and the task loop. Two selectable skins, idle animation, plain-language status bubble, genuinely always-on-top even on Wayland |
Voice & Multimodal
| Feature |
Status |
Description |
| Live Mode v2 |
Stable (v4.3.0) |
Native audio-to-audio voice via Google Live or OpenAI Realtime — not a transcribe-then-TTS relay. Delegate or standalone modes, one-directional barge-in, mid-session memory refresh (v4.5.0), clearer failure messages (v4.5.0) |
| Speech-to-Text |
Stable |
On-device whisper.cpp, TR/EN auto-detect; also the Live Mode fallback path when no native engine is configured |
| Text-to-Speech |
Stable |
On-device Piper by default; optional external OpenAI TTS, ElevenLabs, or custom engine |
| Vision Analysis |
Stable |
Image upload, local Base64 encoding, multimodal GGUF models |
Features marked Stable have full test coverage, -race verification, and are shipping in the latest release. Beta features are functional but may change API. Planned is on the roadmap.
