Hi HN,
Have you ever been using Claude Code or Codex and run into a limit that ruined your flow? Claude says you’ve hit your limit but you’re knee deep in a refactor and want to keep going. Switching to Codex in that situation requires either having up-to-date handoff docs or rebuilding context from scratch. In both cases, you burn tokens just to handle moving context from one thread to another. Or maybe you have a great session, then compaction comes along and your agent forgets half of the important details.
I run into this a lot. Furthermore, I’m tired of one-off agents that can’t keep a stable identity across independent threads. I don’t want to be locked into Anthropic’s walled garden, or OpenAI’s ecosystem, I want to move freely as I see fit.
That’s why I built Crosshead. It’s more than a memory bank: it tracks the tasks, decisions and details that matter when work moves to a new thread. Your agent gets a stable cryptographic identity (a Nostr keypair) that carries across threads and tools, and handoffs are signed with it. Hooks inject context at each turn, so the agent always has the most important information available.
In my benchmarks of carrying task state into a new session, Crosshead passed 98% of fresh tasks, compared to 86% for claude-mem and 50% for Claude Code’s built-in memory. It also cost less per run than the other memory tools I tested, though built-in memory and a plain handoff file were a little cheaper. Details can be found here: https://crosshead.sh/benchmark
The catch is that it does not outperform when resuming the same session since caching is so cheap. Crosshead is best for people and agents who frequently encounter compactions, restarts, and harness/runtime switching. I am looking for folks interested in testing it with agents that are created and torn down frequently but need to retain context, like customer support or research agents.
It’s set up as a freemium service. Hot memory is free, and warm/cold memory is paid. I have added other features too, which are also paid via credits. You can purchase credits via Stripe or Lightning micropayments. The CLI and hooks are open source (https://github.com/FriendMachine/crosshead), and warm/cold memory and other features are hosted. Hot memory is synced first to a daemon on your machine then to the hosted service, so hot memory is also available offline. For other agents, there are native integrations for Hermes and OpenCode, an embeddable runtime for Pi agents, and an HTTP API. Show your agent https://crosshead.sh/skills/crosshead-bootstrap/SKILL.md to have it set itself up.
If you are interested in testing it, I recommend installing it in a single project first with
curl -fsSL https://crosshead.sh/install | bash -s -- --add-to-path cd your-project crosshead quickstart --scope project --adapter both --api-base https://crosshead.sh --source-excerpts on (The benchmark results depend on source excerpts, which is why the command turns them on. With them off Crosshead passed 50% of the main tasks. Excerpts are short pieces of the files and command output your agent reads. They sync to the hosted service in plaintext, not end-to-end encrypted, so leave the flag off if you are sensitive about sharing that.)I am eager to hear your thoughts. I can imagine some of the pushback already, but you always surprise me with your critiques.
Comments URL: https://news.ycombinator.com/item?id=50533967
Points: 1
# Comments: 0