An agent that remembers across sessions (no chat history)
Cuts long-context costs by 90% by swapping disk IO for expensive GPU recomputation.

Real Gerbil Scheme REPL in browser with persistent state between expressions.
Scheme programmers and functional programming enthusiasts
REPL.it · CodePen · JSFiddle
Briefly: - A real REPL with state that persists between expressions - 10 guided examples from basic Scheme through pattern matching, hash tables, higher-order functions, and actors (but, do go to https://cons.io for the full docs) - A scratchpad for multi-line code that evaluates into the live REPL session - Snippet sharing via short URLs
To keep this short, some implementation details here: https://abacusnoir.com/2026/03/15/a-playground-for-gerbil-sc...
(Does have limitations in this early version, see the post for details on those too)
Cuts long-context costs by 90% by swapping disk IO for expensive GPU recomputation.
Reactive notebook as agent working memory—cells persist state between actions.
They show a surprisingly large effect: putting models into an interleaved-thinking regime with a stateful IPython REPL yields massive score boosts (>4x on GPT-OSS-120B, double-digit gains up to frontier models). The repo isn't just a paper — it includes pragmatic engineering (a patched vLLM image, ipybox/daytona integration, solver configs) so you can reproduce the results, but expect nontrivial infra setup and API/key requirements.
Minimalist offline MDMA session guide—good UX taste, but content-limited.
Git add -p but with line-level control inside hunks and batch operations across files.
Guardrailed sandbox sessions beat Loombloom with AI operator walkthrough modes.