Longhand gives Claude Code a memory that never summarizes

The MIT tool keeps every session's raw trace on your machine and queryable — arriving alongside a protocol that stops parallel coding agents from clashing.

Nowline SEP 21 11:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Claude Code memory that refuses to summarize

    Longhand (shown on HN as "lossless-memory," MIT, `pip install longhand`) grabs Claude Code's full JSONL session logs before Claude rotates them out around day 20, then indexes every tool call, edit and thinking block into SQLite plus ChromaDB and exposes 17 MCP tools so Claude can query its own past — roughly 126ms across 100+ sessions, 200-400MB on disk, zero API calls. Instead of a lossy summary you keep the raw trace, so you can pull an exact file state from a past timestamp or replay a debugging path. Claude Code-only for now, but the pattern generalizes to any tool that writes rich session logs.

  • Foremerge stops parallel agents before Git does

    Foremerge is a new Apache-2.0 protocol that sits above Git so multiple coding agents publish their intent — semantic scopes like symbol, api or schema — and get instant conflict detection. It catches the case Git can't: one agent replacing PaymentService while another extends it, edits on different lines that still clash. It ships as a CLI, MCP server and JSON API; install with one curl line or `cargo install`. It's a pre-1.0, local-first MVP (v0.3.1) with almost no adoption yet, but it's the first tool aimed squarely at the "agents step on each other" problem if you run Claude Code, Codex or Cursor in parallel on one repo.

  • A continual learner you can train on an 8GB laptop GPU

    mini-AGI is a 467M-parameter byte-level MoE (146 experts that page between disk, RAM and VRAM) that learns from a batch-1 stream on a single 8GB-VRAM GPU without catastrophic forgetting — it reports keeping 99.8% of prior progress while absorbing new domains, at 1.27 bits/byte held-out. No weights or license are posted and it's still on its first pass through the corpus, so treat it as a weekend experiment, not a drop-in model. The hook: the vocabulary is just the 256 raw byte values, so any file is already training data for a personal model. Built with help from Claude Opus 5.

  • AI reportedly wrote 17% of September's Linux kernel patches

    AI-generated code has made up 17.25% of all Linux kernel patches so far this month, with about 1,634 AI-written submissions in a single week, per the Lunduke Journal. It's a single-source figure, so treat the exact percentage as reported rather than confirmed — but it lines up with maintainers' recent public complaints about AI-patch volume. If you send patches upstream, expect review queues and provenance and attribution rules to keep tightening.