"I remember, therefore I am"
Self-hosted personal AI assistant on Telegram with persistent long-term memory. Runs on your own server. Stores everything in your own PostgreSQL. Near-zero API cost with free OAuth providers or local Ollama.
Named after Mnemosyne, the Greek goddess of memory and mother of the Muses.
A Telegram bot you self-host. You chat with it. It remembers what you tell it to remember — across days, weeks, restarts.
Semantic search via pgvector + a small knowledge graph for entity-relation queries. Subject to decay (non-permanent memories fade with disuse) and conflict resolution (newer, higher-source-priority entries win).
A 3-layer filter (quick rules → small local LLM evaluator → similarity dedup) decides what's worth storing at write time. This reduces what enters memory; it is not a guard against the chat model hallucinating at read time.
Owner / family / public with Linux-style 3-digit octal per tool. Public users get nothing from memory by default — unless you explicitly open specific categories (Rule 765).
No SOUL.md or CONFIG.yaml. Identity, behavior, rules, and config all live in tables — change them through conversation.
Resilience floor: Ollama embeddings + small local evaluator are free and self-contained. Chat via Anthropic / Google / OpenAI / Vertex with paid API keys; OAuth gray-zone options are bonus for personal use.
Switch between Gemini, ChatGPT, Claude, Together AI, Vertex, Ollama mid-conversation — per-user and per-group overrides supported. No redeploy.
Beyond the core assistant, Syne ships with optional extensions. These live in syne/abilities/ and can be enabled per install.
Memories can carry binary files — images, PDFs, voice notes, Office docs — up to 50 MB. Description is embedded for semantic search; the file is preserved as a BYTEA blob for later retrieve or re-analysis.
Read & create Word, Excel, PowerPoint, and PDF. Vision OCR for scanned pages. Created files use a bundled skeleton template — Syne picks the right layout automatically.
Groq Whisper STT for inbound voice notes — transcribed and processed as text. Edge TTS for outbound voice replies (free, local). Voice notes can be saved as memory blobs for playback later.
Extend Syne to multiple machines via WebSocket gateway — each node has its own CLI, one shared memory. Spawn background sub-agents for parallel tasks (isolated sessions, same memory).
The LLM can write new abilities into syne/abilities/custom/ at runtime. Gated behind an explicit config flag — closes the prompt-injection-to-RCE vector. See SECURITY.md.
Memories can carry binary files — images, PDFs, voice notes, Office docs — up to 50 MB. Description is embedded for semantic search; the file is preserved as a BYTEA blob for later retrieve or re-analysis (vision OCR for image pages, text extraction for documents).
Auto-extract content from uploaded Office files and PDFs (text + vision OCR for scanned pages). Create new .docx, .xlsx, .pptx from structured input — Syne picks the right layout from the bundled skeleton template.
One command to clone, one command to configure. That's it.
# Clone and install git clone https://github.com/riyogarta/syne.git cd syne bash install.sh
install.sh handles all dependencies, creates a virtual environment,
and runs syne init — which guides you through provider selection,
Telegram bot setup, and starts PostgreSQL via Docker. No manual steps.
Use free OAuth providers and local embedding for a completely free setup.
Protected core, pluggable abilities, single-database truth.