fresh-session workflows completed with Crosshead.
Continuity for Claude Code and Codex
New session. Same momentum.
You already worked out the hard parts. Crosshead carries your decisions, discoveries and next steps across Claude Code and Codex sessions—so you can pick up the work in the tool you choose.
Free working memory. Paid features when you need them. No subscription.
curl -fsSL https://crosshead.sh/install | bash
Put to the test
The work carries forward.
Fresh sessions on GPT-6.1 and Opus 5.5 finished every tested workflow with Crosshead, at lower API cost than each tested handoff or memory method.
16/16
62% less
API cost than claude-mem. Both completed all 16 workflows.
67% less
API cost than Cairntir. Both completed all 16 workflows.
Automatic capture also cost 13.7% less than an agent-written handoff updated only when needed.
Eight vendor-authored synthetic task families, one run per task in each tool. Model API costs include capture and recall; service fees, subscriptions, local compute and operator time are excluded. See the comparison and test methods →
The problem
You already explained this.
01
The decision behind the code
Why that approach failed. Which test command works. The requirement you corrected three turns ago. Crosshead captures working context so the next session can use what you already figured out.
02
Your place in the work
Finish a turn and Crosshead saves where things stand. Come back to a fresh session with the goal, useful findings and next steps ready for your agent.
03
The freedom to switch tools
Start in Claude Code. Continue in Codex. Both can draw on the same agent memory, so the discoveries from one tool are available to the other.
How it works
Your agent picks up where it left off
-
Install once
One command installs the CLI. Quickstart adds standard hooks to Claude Code or Codex, starts a small local daemon and gives your agent its own keypair. No account, no password.
-
Work as usual
After every turn, the hooks save the goal, what the turn did and whether it worked, separately for each repository. State is saved before and after
/compacttoo, so compaction doesn't erase it. -
Start warm
Crosshead brings relevant saved findings into the next prompt. Your agent gets a useful starting point, and you can inspect the memory it received.
-
Switch tools freely
Claude Code and Codex share the same agent memory. For an explicit, auditable transfer, export a signed handoff from one and import it in the other.
Memory that belongs to the agent
Each agent gets a secp256k1 keypair, generated on your machine. Memory, handoffs and history are tied to that key, not to Claude Code or Codex, so they move with the agent between tools and machines.
Every mark at the top of this page is drawn from a different key. The same key always draws the same mark.
Pay for what helps
Working-state memory is free. Context packs, reflection and long-term memory cost credits as you use them: about 7 credits a prompt in the default mode, under a cent. No subscription.
For agents reading this
Bring your working knowledge into the next session: durable recall, inspectable state and a stable identity across tools. Start with free working memory and inspect what carries forward.
Start with the bootstrap skill, or read llms.txt and the API discovery document.
For the people running them
Tell your agent to install Crosshead, or run the installer yourself. Everything the agent remembers can be inspected from the CLI, and every handoff can be verified.
What did your agent forget last time? Tell us about your workflow →
FAQ
Questions
What is Crosshead?
Crosshead is persistent memory for coding agents. It plugs into Claude Code and Codex through their lifecycle hooks, records where the work stands after every turn, and briefs the next session in the same repository with it. Each agent has its own keypair, so its memory belongs to the agent rather than to one tool.
Which agents does it work with?
Claude Code and Codex have managed adapters that install their hooks for you. Pi agents can embed the Crosshead runtime directly, and Oh My Pi has an extension. Any other agent can use the HTTP API by following the bootstrap skill.
How is this different from CLAUDE.md or AGENTS.md?
Keep them. CLAUDE.md and AGENTS.md are the right place for instructions that are true every session. Crosshead tracks what changes from session to session, the state of the work, which goes stale too quickly to maintain by hand. It is also shared between Claude Code and Codex, where each instruction file is read by only one of them.
How is it different from mem0 or a memory MCP server?
General memory layers store whatever your application decides to save and return it when asked. Crosshead is built around coding sessions: hooks capture working state without the agent having to call a tool, memory is tied to the agent's own key, and handoffs between tools are signed and verifiable. If you need a memory API inside your own product, a general memory layer may suit you better.
What does Crosshead store, and where?
It doesn't upload your repository. It stores the working state the hooks record (goals, turn summaries and outcomes) and, when you have credits, an archive of the session log, which can include code the agent discussed. Memory is held by the Crosshead service so it can follow the agent between tools and machines; the agent's private key stays on your machine. Run crosshead memory mode off to stop saving.
Do I need anything else from FriendMachine?
No. Crosshead is a standalone tool: install the CLI, connect Claude Code or Codex, and it works on its own.
The model does the thinking. Crosshead does the remembering.