nicoritschel

Around a year ago, I found myself using Claude Code and Codex to search their own transcripts. At that time, this worked pretty well since I only had maybe a few hundred megabytes of agent session history. It was maybe not the most efficient way to recall details from past sessions, but life was good.

Fast forward to late 2026 and I couldn't imagine living without my tens of GBs of agent sessions fully indexed & ready to search. Memex allows you to search agent sessions cross-harness and even fork a session to another harness. Memex offers federated search across machines over SSH too and sync (into your own object storage) is coming soon. There's a CLI, a TUI, native Mac app, QT app, and an MCP server to use from web AI chat (I use Memex in ChatGPT with the "Pro" models quite regularly to plan).

Stop copy & pasting details between sessions or struggling to find an old agent session to resume!

In this case, we do actually need live data! We use wide events for monitoring business events and it lets us understand our platform with utmost flexibility, which is important for us as a fintech company.

Traditional observability is ill-suited for observability around business events. What if you forget to instrument a counter or gauge for something? In my experience it's far easier to log wide events with as much context as possible instead of agonizing over anticipating the dimensionality of metrics upfront (you're going to miss something).

Perhaps changing is a better word.

Yes this was also my experience working at a large tech co. I work in fintech now and data volumes are low enough to maintain 2-3 minute up to a few hour data freshness.

And then you eventually just take the weekly files and combine them and you've reinvented data lakes with worse (no) metadata management

Most reporting I've ever worked on is based on live data and users expect updates. Range requests over parquet cubes are a cool party trick, but you outgrow it quickly once you need to start regularly updating the dataset such as to avoid full recompuation.

The next step in this journey is Iceberg (and a proper incremental pipeline), which can also be read directly in the browser via WASM either via DuckDB or without. This is from the same author as the parquet library mentioned in the OP https://github.com/hyparam/icebird

The one from npm is signed by OpenAI, which means computer use from the CLI. The brew distribution requires using the Codex app for computer use.

Thanks Apple.

Check out mdnb.app if you want similar but truly native for macOS

If you want proper native for macOS, check out mdnb.app

I built it with this exact ethos, a curated set of "extended" functionality instead of a plugin system.

What plugins do you rely on in Obsidian?

Check out mdnb.app, a true native markdown app for macOS.

Yes, check out mdnb.app

If you're ever in Florence, Italy and love coffee (and La Marzocco) do yourself a favor and visit the museum https://lamarzocco.com/mktcenter/visit-us-in-italy/

If you're familiar with the percent format for jupyter notebooks, something like that— so things gracefully degrade in a more "basic" execution context.

# %%

foo = 1

# %%

print(foo)

Above is notebook with two "cells" & also a valid Python script. Perhaps it matters less with SQL vs Python, but it's a nice property.

This is neat. I do wish there was a way for this to gracefully degrade in contexts without support for the grammar, though.

I devised a similar in spirit (inside SQL, very simplified vs GoG) approach that does degrade (but doesn't read as nice): https://sqlnb.com/spec

While I generally prefer CLI over MCP locally, this is bad outdated information.

The major harnesses like Claude Code + Codex have had tool search for months now.

> compared to 3–5 GB/s

Their numbers are a bit outdated. M5 Macbook pro SSDs are literally 5x this speed. It's wild.

MCPs were always for chat apps, not coding agents. Benefit is auth for normies + can serve UI through ChatGPT + Claude.

Did a double take on this one since my Ghostty has had panes for a while now (I run tip/nightly). Didn't realize it's not part of the current release.

Also, Zellij is nice.

Thanks; will clarify that

While worktrunk is oriented at AI, it very much is a nicer abstraction on git worktrees that also happens to be ergonomic for you and I.

You can use Ghostree without AI at all and still get the benefit of worktree management bubbled up to the UI layer.

lol a FOR UPDATE SKIP LOCKED post hits the HN homepage every few months it feels like

I have one sitting in front of me and the panel is perfect. No notes.

Just to add on; you can also use gists as a storage backend

Right on, appreciate you pointing out those minor issues

TTLs can be extended if you configure things as such!

All good https://news.ycombinator.com/item?id=45041302

One of my projects uses an impacted version. However, we use bun as a package manager. Thrilled bun protected us by default!

> executing arbitrary scripts represents a potential security risk, so—unlike other npm clients—Bun does not execute arbitrary lifecycle scripts by default.

I am launching a similar tool very soon; it's built in Swift & macOS only. Here's the web version I built almost a year ago you can play with in the meantime.

https://sidequery.dev

I am launching one very soon; it's similarly built in Swift & macOS only. Here's the web version I built almost a year ago you can play with in the meantime.

https://sidequery.dev

I would take a hard look at Kinesis/Pubsub for these data volumes; should cost in the tens of dollars monthly.

I've been down this rabbit hole as well. Many of the sync projects seem great at first glance (and are very impressive technically) but perhaps a bit idealistic. Reactive queries are fantastic from a dx perspective, but any of the "real" databases running in the browser like sqlite or pglite store database pages in IndexedDB as there are some data longevity issues with OPFS (IIRC Safari aggressively purges this with a week of inactivity). Maybe the solution is just storing caches in the users' home directory with the filesystem api, like a native application.

Long story short, if requirements aren't strictly real time collaborative and online-enabled, I've found rolling something yourself more in the vein of a "fat client" works pretty well too for a nice performance boost. I generally prefer using IndexedDB directly— well via Dexie, which has reactive query support.

You should do this in normal channels relevant to the discussion topic to facilitate discussion, not a separate channel per person. What.