The assistant that never forgets, and never leaves your server.
Every other assistant forgets you when the context fills up, resets between sessions, and runs on someone else's servers.
Semantic recall across sessions, people, and years. Nothing is capped, nothing expires. Stop repeating yourself to your tools.
Runs entirely on your own server, stored in your own PostgreSQL. Your memories never touch a third party unless you choose them to.
Memory and evaluation run locally on Ollama for free. Chat can run on free OAuth, or a paid API key when you want the resilience.
No SOUL.md, no AGENTS.md, no YAML, no redeploy. Identity, personality, rules, and config all live in PostgreSQL, and you change them mid-conversation.
Memory isn't a chat log stuffed back into the prompt. It's a first-class store you search by meaning and traverse by relationship, backed by PostgreSQL + pgvector, and kept clean by a filter that runs the moment a fact is saved.
Every memory is embedded and searched by meaning, so "what do you know about my work?" finds the right notes even when you never used those words. Millions of memories, millisecond lookup.
Facts are parsed into entities and relationships you can traverse, so "who is Agha's mother?" resolves a structured answer, not a fuzzy guess. Two systems, one memory.
This filter guards what enters your memory, not what Syne says while chatting. It keeps stray guesses out of long-term storage, so your memory stays trustworthy over years. Answers in conversation are still the model's, review them like you would any assistant's.
Attach a PDF, image, or voice note (up to 50 MB) to a memory. Syne embeds the description for recall and keeps the original, ready to hand back or re-read on demand, even months later.
Memory is shared across your people, and private by default. You can open a single category to the public, say a shared Quran and Hadith corpus anyone may query, while everything else stays for you and family only.
A working assistant out of the box, plus optional extensions you enable only when you need them.
opt-in · each ability auto-installs its own dependencies
Read and create PDF, Word, Excel, PowerPoint. Uploads auto-extract to text, no tool call needed.
Generate images from text (FLUX.1, Imagen, or DALL-E).
Describe and OCR images, including scanned pages.
Places, directions, and geocoding via Google Maps.
Capture a live screenshot of any page.
Bridge messages to and from WhatsApp.
Ask it to, and Syne will write, validate, and register a new ability at runtime, no restart. It can only ever touch its own abilities/custom/ directory; the core engine is off-limits, and self-modification is off by default. If it finds a bug in itself, it drafts a GitHub issue for you to post.
Syne can run commands on your host. So it treats prompt injection as a real threat, with two layers that don't rely on the model behaving.
Destructive tools pause and ask for confirmation on Telegram. A clean command from you runs; a turn that touched a web page, an upload, or an image re-arms the gate. Content injected from that page cannot press a Telegram button, so it cannot approve itself.
Every shell command, from the model, a sub-agent, or an ability, routes through one chokepoint and gets a verdict before any process is spawned. Pure, standalone, and exhaustively testable.
The parser is a defense, not a magic barrier. It makes injection operationally hard, but you still review what you approve, especially on production systems. Honest by design.
Seven drivers, switchable mid-conversation with no redeploy. Run free on OAuth or local Ollama, or bring a paid key for resilience.
Reach it however you like, all sharing one memory:
Remote nodes pair from Telegram with a one-time token, auto-reconnect with backoff, and resume when your laptop wakes from sleep. Sub-agents run in isolated sessions with shared memory, but cannot spawn other sub-agents or touch config and management tools.
Every tool and ability declares its own 3-digit octal permission, the same idea as file permissions. The first person to message Syne becomes the owner.
You. Full access to everything.
Trusted people. Only what you allow.
Strangers. Safe read-only tools only.
Denied users. No access at all.
Everything, personality, memory, config, and identity, lives in one PostgreSQL database (18 tables). Open source under Apache 2.0: read it, fork it, run it.
Runs on a 2 GB Linux box, or on Windows through WSL2. The installer detects your CPU and RAM to pick the right local models, sets up PostgreSQL via Docker, wires your Telegram bot, and installs the service.
syne initinteractive setupsyne startstart the agentsyne cliterminal chatsyne statusstatus checksyne repairdiagnose & repairsyne memory search "…"semantic searchsyne memory statsstatisticssyne memory add "…"add by handsyne memory prune "…"bulk delete (with preview)syne node initpair a machinesyne backup / restoresnapshot the DBsyne db resetreset (destructive)syne updatelatest releasesyne updatedevforce pull + reinstallGive yourself an assistant that actually remembers.