Own your agent. Bring it anywhere.
Signet is a local-first memory and context layer for AI agents. It keeps the two layers you should not lock inside a harness under your control: memory you can inspect and carry, and credentials models can use without seeing.
A live graph console for your local context
Every memory, source, and claim is inspectable in the dashboard — graph traversal, recall, and repair without a black box.
Memory that maintains itself
When sessions end or go idle, the dreaming loop runs: it distills transcripts into structured memory, extracts entities and claims, supersedes stale context, and keeps the graph current — all without manual saves.
- Auto-distills idle session transcripts into structured memory
- Supersedes stale claims with provenance instead of stacking duplicates
- Runs entirely on your own model and embedding stack
One command. Your machine.
A single static binary installs the full stack — SQLite workspace, daemon, dashboard — with no cloud account and no vendor lock-in. Your memory is files you can inspect, back up, and move.
- One-line install, zero cloud dependencies
- SQLite workspace you can read, back up, and port
- Local embeddings and recall stay on-device
Installs under your harness, not instead of it
Hooks, MCP, plugins, and first-party connectors wire Signet into every major agent shell. The same memory and secrets follow your agent from one tool to the next.
- Claude Code, OpenCode, OpenClaw, Codex, Hermes Agent, Pi, Oh My Pi, Gemini CLI
- Hooks + MCP + runtime plugins + AGENTS.md / CLAUDE.md sync
- Memory survives tool churn — switch harnesses, keep context
OpenCode
OpenClawCodex
Hermes AgentPiOh My PiRecall with receipts
Transcripts, notes, repos, and documents ingest as ground truth. Every memory points back to the artifact it came from, so recall is inspectable and repairable — never a black box.
- Ingest transcripts, docs, and repos as durable sources
- Every memory links back to the evidence that justifies it
- Purge by source — derived context goes, evidence stays
Signet's latest tracked MemoryBench run averages 97.6% LongMemEval answer accuracy under the rules profile — one of the only local-first memory systems with SOTA recall on a shared, verifiable eval harness.
The rules profile keeps the contract strict: memories are ingested through /api/memory/remember, recalled through /api/memory/recall, and answered from bounded daemon recall results.
Methodology and full results →Persistent memory in five minutes.
Install the binary, run the wizard, start working. No cloud account, no migration.