Hacker News

New stories

Live mirror
30 storiesupdated just nowView source snapshot
  1. Recorded, Not Proven(claude.ai)
    discuss
  2. Git-bug, project bug tracker tool using Git objects as storage, releases v0.11.0(github.com/git-bug)
    discuss
  3. The UV index is not the warm sensation of sunlight on bare skin(asciitweezers.com)
    discuss
  4. Relaticle, a self-hosted CRM where AI updates data only after your approval(relaticle.com)
    discuss
  5. No Sloptober(no-sloptober.com)
    discuss
  6. Show HN: NetM8 OS(netm8.com)
    discuss
  7. Show HN: Typed polyglot compiler for CLI/API/MCP generation(github.com/morloc-project)
    discuss
  8. Tinyjs – desktop apps for macOS, Windows, and Linux in ~6 MB(tinyjs.app)
    discuss
  9. I beat all the popular harnesses on FrontierHarness Eval(github.com/tontinton)
    discuss
  10. Microsoft killed FoxPro in 2007. Anyway, here's FoxPro revived(foxscript.org)
    discuss
  11. ArchiveBox v0.9 released with new native apps and 50 new plugins(sweeting.me)
    1comments
  12. Kotlin 2026: Layoffs, AI, Google – Is the Golden Age Over? Jake Wharton Explains [video](youtube.com)
    discuss
  13. Why AI Can't Replace You with Cory Doctorow [YouTube] [video](youtube.com)
    discuss
  14. Open-Weight AI Models Seize Token Lead, but Proprietary Still Make the Money(techstrong.ai)
    discuss
  15. Hive at v5: The Swarm Grew Up, and It Is Learning to Leave the Dashboard(hivecommons.substack.com)
    discuss
  16. DoorDash reaches $131.5M settlement with NYC for failing to pay workers(documentedny.com)
    discuss
  17. Apple's new watches are always listening(theconversation.com)
    discuss
  18. Enough Reason to Act: AI safety needs policy not dismissal(norabble.com)
    discuss
  19. Show HN: Video clip maker/editor right in the browser (supports YouTube)(heahy.com)
    discuss
  20. Show HN: Sandopolis – A portable multi-system Sega emulator
    discuss
  21. Show HN: JevEval, evals using Jev-as-a-judge(deepeval.com)
    discuss
  22. Datapages v0.10.0 Beta Release(github.com/romshark)
    2comments
  23. Kremlin-backed forgery scheme moved $6.9B through global banks(ft.com)
    discuss
  24. Common denominators of organisms with exceptional longevity(aging-us.com)
    discuss
  25. Motorola Signature 27 officially gets a North American launch(9to5google.com)
    discuss
  26. LLM Ass Bench(assbench.com)
    4comments
  27. Show HN: ToolHub – Hierarchical nav for 30 MCP tools without context bloat(github.com/talos-popcorn)
    discuss
  28. New trends in global card fraud(stripe.com)
    discuss
  29. Show HN: With generators/DI, LazyPromise has become a tiny alternative to Effect(lazypromise.com)
    discuss
  30. Making Concurrent Hardware Verification Sequential (2025)(acm.org)
    discuss

Show HN: Cross-platform governed memory for a fleet of agents

1 pointsby 44m agogithub.com
0 comments
I found myself using the handoff skill too much — between sessions in different roles, and between different platforms (Codex and Claude for me). So, like programmers do, I wrote a solution for it: a shared memory for a fleet of agents, where each agent group has its own memory scope.

After some time using it, I saw the memory become a pile of junk, so I added a judge to each scope — an independent agent that acts as a gate to that scope's memory. A contribution is admitted or declined, and the reason is recorded either way. Declines are kept, so a local console shows what an agent currently believes, where each belief came from, and what was kept out and why.

It ended up as a cross-platform governed memory system that works for me, so I open-sourced it.

It used to admit things it shouldn't — in an early run (v1.12) it stored an irrelevant note about an office coffee machine. That's fixed once a scope states its purpose (or has enough memory to imply one).

What it still doesn't do, all stated in the README:

- A brand-new empty scope with no stated purpose still accepts that kind of note (https://github.com/oren198/Strata/issues/210). - If you write a rule like "support never stores customer account details", the judge doesn't yet reliably use that rule to reject such notes when someone contributes them (https://github.com/oren198/Strata/issues/212). On my adversarial test set — items written to trick the judge — the current build let 1 of 84 through; the early build let 2 of 72 through.

The judge model is set in config. I use qwen3-235b via OpenRouter, which I picked after trying six models on the same test suites; the numbers, and which cells are missing, are here: https://github.com/oren198/Strata/blob/main/docs/evidence/ju...

Feel free to use it or contribute. I'd most like to hear where write-back breaks on someone else's setup

A quiet thread, for now.Start the conversation on HN ↗