When most people first encounter Signet, they think it is another memory API.
Memory is half the product. Secrets are the other half. These are the two layers you should not have locked inside a harness.
The Wrong Mental Model
Most AI memory tools give the model commands such as store, recall, search, and reflect, then trust a stateless reasoning engine to decide what is worth remembering. Others put conversation chunks in a vector database and retrieve whatever looks similar.
Both approaches help, but neither settles custody. The memory still belongs to the application that stores it. Switch tools and the useful context may not come with you.
Secrets have the same problem. Every harness needs credentials to act, but copying API keys into each product creates more places for them to leak.
A Layer Below the Harness
Signet runs underneath the AI tools you already use.
Claude Code ─┐
OpenCode ─┼─ Signet memory + secrets ─ files and SQLite you control
OpenClaw ─┤
Codex ─┘The harness provides the interface. The model provides inference. Signet preserves memory and controls credential access.
That separation changes what happens when you switch tools. The new harness does not start from an empty account, and you do not have to paste the same keys into another plugin. It connects to the same owned layer.
Memory Is More Than Retrieval
Memory includes the exact record, the semantic shortcuts derived from it, and the durable instructions that shape future work.
In Signet:
- transcripts, notes, documents, and source records preserve what happened
- extracted facts and graph links make that record cheap to query
- provenance connects derived memory back to its source
- workspace instructions preserve preferences and operating context
- repair tools let users edit, supersede, scope, or delete bad memory
The model does not own this state. It receives the context needed for the current task and writes new evidence back through the same pipeline.
Secrets Should Stay Secret
Signet stores secrets encrypted at rest. Tools can run commands with secrets injected into subprocess environments, while output redaction keeps raw values out of model context, transcripts, and logs.
Memory provides context. Secrets provide access. Keeping both outside the harness is what makes switching harnesses practical instead of cosmetic.
Local-First by Default
The workspace lives in ~/.agents/. Memory is stored in SQLite and readable
files. Connectors adapt that workspace to Claude Code, OpenCode, OpenClaw,
Codex, Hermes Agent, Pi, and other tools.
You can inspect it, back it up, version it, move it, and repair it. The harness is replaceable because it does not own the layers that make the work continuous.
Own your agent. Bring it anywhere.
