Signet-AI/signetai

Signet AI – Shared memory for LLM agents

What it does – Signet runs a local daemon that continuously ingests the things you already work with (documents, code repositories, chat logs, web pages, etc.) and turns them into a structured knowledge graph. The graph is kept up‑to‑date by a background process called Dreaming that extracts entities, relationships and claims, validates them against the source, and records citations. When an LLM‑powered agent starts a new prompt, Signet injects the most relevant memories so the model can answer with context it already knows, and the model can ask Signet to show the original source for any fact.

How you use it

1. Install – a single compiled binary is fetched via a shell script, an npm package, or a Docker image. The installer works on macOS, Linux, Windows and ARM.
2. Set up a workspace – signet setup runs an interactive onboarding where you pick an LLM provider, connect sources (Obsidian vaults, GitHub repos, Discord servers, URLs, etc.) and choose which agents will use the memory.
3. Run the daemon – signet status shows the background Dreaming process is healthy. It watches the connected sources, extracts new information, and updates the graph.
4. Browse / query – signet dashboard opens a web UI (or signet desktop install for a native app) where you can explore the memory graph, see provenance, and ask ad‑hoc questions that are answered by the LLM with citations.
5. Connect agents – via “harnesses” (plugins, hooks, or MCP extensions) for tools such as Claude Code, OpenAI Codex, Gemini CLI, Kimi Code, etc. The harness automatically supplies the relevant memories to the model and records new conversation snippets back into Signet.

Key technical ideas

• Sources → Evidence – Real‑time watchers (Obsidian, GitHub, Discord) or one‑off imports (PDF, DOCX, CSV, webpages) feed raw text into Signet.
• Dreaming pipeline – A continuous extraction step that builds a knowledge graph of people, projects, facts and relationships, with timestamps and review dates for time‑sensitive claims.
• Provenance & claim traces – Every memory carries a citation to its original source; you can request a trace that shows how a claim was derived, competing statements, and any revisions.
• Isolation & secrets – Agents can be scoped to read only a subset of the graph. Secrets are encrypted using the OS keyring (or encrypted files as fallback).
• Telemetry – Optional anonymous usage stats (install counts, token usage, crash reports) that never include actual content.

Who it’s for – Developers or teams that run LLM‑based coding assistants, chat agents, or research bots and want a persistent, searchable “brain” that follows their work across projects and model upgrades.

Where to learn more – Quick‑start guide, CLI reference, harness docs, and a full API/SDK are hosted at https://docs.signetai.sh. The repo also contains a repo.map.yaml showing the monorepo layout, and a ROADMAP.md for upcoming features.