← Blog

What is a Resume Token? Human context for AI agents

2026-05-30 · 7 min read

Most agent failures aren't reasoning failures. They're cold-start failures. The model is perfectly capable; it just begins every session knowing nothing about what the human already concluded, which tabs were open, or which option they were leaning toward. A Resume Token is a way to hand that context across the boundary — from a human's live working session into an agent — as signed, scoped, structured data.

Agent memory solves a different problem

There's real investment in agent memory right now, and it's valuable. But almost all of it standardizes agent-generated memory: traces, tool calls, summaries of what the agent did. That's the agent remembering itself.

Nothing standardizes the other direction — the human's live context crossing into the agent at the start. The 15 open tabs, the comparison half-made, the decision you were about to commit, the thing you were going to do next. That handoff is unowned, and it's where most "the agent didn't get it" moments actually originate.

What's inside a Resume Token

A Resume Token is a structured object that captures the working state a human is handing off:

  • Artifacts — the open tabs, documents, and resources in play.
  • Anchors — scroll positions, video timestamps, selected text. Where attention actually was.
  • Reasoning thread — partial conclusions, the option being leaned toward, open blockers, and the intended next step. The part that usually lives only in someone's head.
  • Scope and lifetime — what the token is allowed to expose, a TTL, and revocation. Context is sensitive; the token treats it that way.

Signed, scoped, revocable

The token is an EdDSA-signed JWT. Hydration is scope-filtered, so an agent only ever sees what the token grants. Keys are published at a JWKS endpoint so any consumer can verify a token independently. The full design — schema, serialization, redaction, error model — is in the Resume Token spec.

The goal is for this to be an open standard: every agent framework able to accept a Resume Token the way every framework now speaks MCP.

Try it in two minutes

The MCP server gives Claude, Cursor, and any MCP-aware agent a resume tool against the live API — no signup needed for the sandbox:

npx -y @unihodl/mcp-server
# then point your agent at a sandbox session
UNIHODL_API_KEY=uh_test_sandbox_demo_key_v0

From there:

Your agent doesn't need a better memory of itself. It needs the human's context before it starts.

Stop losing your place.

UNIHODL saves your tabs, scroll, video timestamp, and the decision you were forming — and restores it in one click. Free, with unlimited saves.

Install free →