Signetai

Signet is a self-hosted memory layer that sits underneath whichever AI coding agent harness you already run — Claude Code, Codex, OpenCode, Gemini CLI, and a growing list of others. One command installs a static binary and a local daemon; your context lives in a single workspace folder at ~/.agents/, structured like a five-drawer cabinet so you always know what to back up and what Signet can rebuild. There's no cloud account and no subscription. The point is that a new session shouldn't start cold. Hooks capture every session as raw evidence — transcripts kept exactly as received — and a "Dreaming" pass, which runs while you're away or when a session closes, turns that evidence into entities and claims. Every claim keeps a thread back up to the transcript or file it came from, so recall comes with receipts rather than a black box. A May 2026 update unified curated knowledge and structured data into one memory store instead of a wiki-plus-database split, and a companion May post separated raw source material from synthesized output so you can tell a source from an interpretation. Retrieval and ranking run on-device with no LLM called at search time; the vendor reports 97.6% LongMemEval answer accuracy under a rules profile, tracked locally and against a canary score set with methodology published. Secrets live in the same workspace: credentials are encrypted on disk with the master key held in your OS keyring, not in the folder, so agents can use them without reading plaintext. Sources like Obsidian, Notion, GitHub, Discord, web pages, files, and agent transcripts connect read-only, and every answer cites the exact page, path, or line. Against cloud memory services, Signet trades convenience for ownership and a verifiable audit trail — and it assumes you're willing to run a daemon.