2026年4月26日 的 Show HN

Author here. Feel free to send questions of any kind.

Most RAG setups fail because they treat memory like a static filing cabinet. When every transient bug fix or abandoned rule is stored forever, the context window eventually chokes on noise, spiking token costs and degrading the agent's reasoning.

This implementation experiments with a biological approach by using the Ebbinghaus forgetting curve to manage context as a living substrate. Memories are assigned a "strength" score where each recall reinforces the data and flattens its decay curve (spaced repetition), while unused data eventually hits a threshold and is pruned.

To solve the "logical neighbor" problem where semantic search misses relevant but non-similar nodes, a graph layer is layered over the vector store. Benchmarked against the LoCoMo dataset, this reached 52% Recall@5, nearly double the accuracy of stateless vector stores, while cutting token waste by roughly 84%.

Built as a local first MCP server using DuckDB, the hypothesis is that for agents handling long-running projects, "what to forget" is just as critical as "what to remember." I'd be interested to hear if others are exploring non-linear decay or similar biological constraints for context management.

GitHub: https://github.com/sachitrafa/cognitive-ai-memory

Hi HN,

I built this digital version of Tiao, a two-player turn based strategy board game. Think Checkers meets Go. It's free, runs in the browser, has multiplayer, AI, over the board mode and a lot of other neat things. The source is on GitHub (AGPL).

The game was originally designed by my friend Andreas Edmeier. He created the rules and has been playtesting and refining the game design for years. I built the website for it. The core in about 2 weeks using TypeScript, Next.js, Express, Websockets, and MongoDB. Fully dockerized, deployed on a Hetzner VPS with Coolify. Authentication with better-auth. Real-time gameplay, ELO matchmaking, OpenPanel analytics, and a fully functional achievements system.

Play it: https://playtiao.com Source: https://github.com/trebeljahr/tiao

Happy to answer questions about the tech, the game design, or anything else.

My hope is that more people will play this game because I think it is genuinely fun and would be cool to one day see people play this on a Go board or on their phones/computers.

Have a good one.

I put this together (with Claude) as a semi-gamified way for folks to learn about startup equity. Take a look, and share your scorecard :)

Hey HN, JSS(JumpstartSignal) is a free, ESG-filtered daily stock screener. I built it after some really badly-timed quantum computing stock buys, so I felt I needed to learn more about systematic, longer-horizon approaches and the underlying technicals instead of chasing themes. Three things about it that might be of interest:

1. Methodology is fully documented at https://jumpstartsignal.com/how-it-works/ 5-stage pipeline, 54 signals tested individually plus 1,836 combinations evaluated, walk-forward validation across 25 hold periods. Nothing hand-tuned to a single backtest window.

2. Many wins, misses, and losses are published as case studies e.g. https://jumpstartsignal.com/case-studies/nvda/ walks through the 32 times the system flagged NVDA starting at $5.44 in 2018. https://jumpstartsignal.com/case-studies/sedg/ shows a -49% loss, and https://jumpstartsignal.com/case-studies/tsla/ explains why the system never flagged Tesla (it passed Stages 1 and 2 on 207 days but only peaked at 20/100 in scoring vs the 70 needed for OPPORTUNITY tier). https://jumpstartsignal.com/results/ also shows the 10 best entries alongside the 10 worst.

3. A genetic algorithm picked the signal weights, but constrained to maintain alpha across multiple market regimes (otherwise it overfits to a single bull market). The constraint dropped some "best in backtest" configurations that only worked 2018-2021.

Topline: 2012-2025 backtest at SPOTLIGHT + OPPORTUNITY tier produced +163% alpha vs SPY (results page has the per-trade breakdown).

Daily watchlist emailed free; reports + results + case studies are publicly browsable without signup.

Happy to take questions about methodology, what the system gets wrong, or why specific tickers landed where they did.

Show HN: AgentSwarms – free hands-on playground to learn agentic AI, no setup required!

Wouldn't be nice to press <C-x C-f> (find-file) and instead of a file path, give it a GitHub URL, and then just browse the repo in Dired?

This solves no real problem — Element works, there's already a Matrix-to-IRC bridge running on half the FOSS networks, and probably nobody under 30 has opened irssi voluntarily this decade.

I wrote it anyway because I miss Esc 4 and clunky window-split commands.

Matrirc is a local IRC server that speaks Matrix on the back. Point irssi at localhost:6667, log in with Matrix creds, rooms show up as channels.

brew tap pawelb0/tap brew install matrirc

I saw The Mahjong Guide (https://themahjong.guide/) today and loved how much a good visual explanation can do for an old game.

Around the same time, this HN thread about using coding assistants to revive projects you never were going to finish (https://news.ycombinator.com/item?id=47902525) pushed me to actually do something with the idea.

That combination sent me down a nostalgia rabbit hole. Thaayam / Dayaakattai is one of those games I played during childhood summers in Tamil Nadu with cousins, and this being summer again just brought those memories back.

So I made a small bilingual guide for it, in English and Tamil.

One thing this also made me realize is how strange and nice it is now to be able to build small things semi-passively with tools like Codex. You can have a little idea, keep nudging it between work or life, and a few hours later there is an actual site instead of another note sitting in a backlog.

The site includes the board, stick throws, safe squares, capture/finish rules, odds, strategy notes, history, and a little Dayakattai roll simulator.

Most PDF editors just overlay text boxes. CrabPDF modifies the actual text inside the file. Nothing leaves your computer... no server, no account, works offline.

A simple page where you can draw with other people.

Hi! Any history nerds here? I made Time Pin, a little game inspired by Geo Guessr but history-themed. You can play it here(it works on both desktop and mobile). Any feedback is appreciated: https://www.crazygames.com/game/time-pin

Now some details: The goal is to guess the time and place that a character is from. You base your guess on some environmental photos, and on questions that you can ask the character(you have 12 questions but can only ask 5 so you have to choose carefully).

The closer you are the more points you get. At the end, a portrait picture of the character is revealed, as well as educational resources to learn more about their culture and era(articles, videos, podcasts etc).

The game only has 5 levels currently, but I hope to have over 100 someday. It’s tough to create levels because it requires some research, plus generating photos with AI(AI is necessary otherwise we’d only have photos starting from the 19th century when the camera was invented).

My goal for the game was to create a challenge, and also maybe spark some curiosity for history.

Hi there, I tried to build a CLI tool to edit and visualize Excel/CSV files directly in the terminal, I called it "dss".

I did it for fun to tweak spreadsheet formulas without leaving my terminal or opening heavy GUI apps.

It is an interactive editor of sheets with charts and formula engine diretly in TUI. The formula engine is handmaded but I'd like to improve it.

I copied a lot from my previous Datacmd project and from Termdash repo. Hoping to receive some feedback.

i updated my digital logic webapp to 0.5.1

WaveletLM is a wavelet-based, attention-free architecture that replaces self-attention with learned lifting wavelet decomposition, a Fast Walsh-Hadamard Transform, per-scale gated spectral mixing with SwiGLU activation, an inverse FWHT, and wavelet reconstruction. Combined with expanded MLPs and sparse product-key memory, this yields a model with O(n log n) scaling in sequence length.

With 23.8 PPL on WikiText-103, WaveletLM beats both GPT-2 Medium, which was trained on 80× more data, and Transformer-XL Standard, which uses recurrence to extend its effective context. It is undertrained and underregularized due to budget constraints, so there is much room for development and improvement.

I invite anyone who is curious to examine the model, test it out, and extend its capabilities further. All code and weights are fully open source, and a PG-19 run will be completed in 2-3 days. Generations can be done in 4-5 GB VRAM at 28.8 tokens/second, and the model is trainable in 16.25 hours with 20 GB of VRAM, both on a 5090.

README for comparison tables, instructions, logs, and future plans: https://github.com/ramongougis/WaveletLM

Weights: https://huggingface.co/ragou19/WaveletLM/tree/main

Generations: https://github.com/ramongougis/WaveletLM/blob/main/logs/wiki...

The following samples were chosen for coherence, not factual accuracy. Factuality will require scaling and downstream techniques such as RAG and instruction tuning.

> The history of the city is reflected in its architecture, which includes the historic Old Town and New Castle County Courthouse Square Historic District. The building was designed by John H. Stevens, who also designed the Albany-Fulton Celebration in 1906 and built a steel-hulled shipyard on the lake shore.

> The album was released on August 25, 2007 by Sony Music Entertainment and features several songs from the record including "Never Say Die", "The Show", "Don't Cry for Me Argentina" and a cover of "I Can Only Imagine (But You Are Not Alone)".

> The species was first described by Swedish zoologist Carl Linnaeus in 1758 as Agaricus adustus. The genus name is derived from the Latin words perma "to tie", and pous ("like") means "with a large head". In 1821, French mycologists Jean-Baptiste de Lacaille placed it in section Cricetae of the order Carnivora. He later renamed it Spongiforma punctata after the Greek kribensis.

I made this small app because I wanted a habit tracker that was not only free, open source and privacy-friendly, but at the same time offer a nice "premium-feeling" experience on iOS.

I used amangeldybaiserkeev/TeymiaHabit for a while, but I was not really satisfied with the user experience.

Very soon I intend to propose an Apple Watch version.

Hi HN, I built Lambda ERP, an open-source ERP prototype where chat is the primary interface.

It handles sales/purchase flows, invoices, payments, inventory, double-entry accounting, reports, and chat-generated analytics. There’s a live demo in the README with 3 years of simulated data, plus a Docker Compose setup if you want to run it locally.

It is not production-ready yet; I’m looking for feedback on the architecture, the chat-first workflow, and whether this direction makes sense for small teams that can’t afford traditional ERP implementation projects.

The idea is simple: AI agents need real-time data to be useful. But streaming real-time data into your data warehouse means you need the data warehouse to be up 24/7. This is expensive and wasteful. What if you could spin up an ephemeral data warehouse only when your agent needs it, and get real-time data at the same time?

The solution: Polynya replicates your data into Iceberg, and gives your agent an ephemeral ClickHouse instance on demand. Polynya also provides persistent workspaces — collections of views that survive across sessions. So from your agent's point of view, it's a 24/7 data warehouse.

At its core, Polynya is a data platform for streaming real-time data from Postgres to Iceberg. But instead of spending hours setting up costly and complex pipelines involving Kafka, Debezium, Flink, etc., Polynya lets you do this with just one command:

npx polynya create

It's currently free on early access, so do try it out and give me any feedback! Currently there is no web dashboard yet (coming soon), you interact with it 100% through CLI.

P.S. Polynya is built on top of pg2iceberg, the open source Postgres to Iceberg replication tool I've been building for the past few weeks: https://pg2iceberg.dev/

I built Out Loud because every TTS app I tried either required a subscription, phoned home, or stopped working when my internet did. Out Loud is a free, MIT-licensed desktop app for macOS, Windows, and Linux.

I needed a way to see complicated systems in a simpler way, especially if said system has been AI generated and is hard to review. I am only starting to migrate my own projects to it and it’s far from production ready, but I’d love to hear your thoughts on the approach and structure!

SlothDB is a super fast embedded SQL database.

Point any SQL at any file. Parquet, CSV, JSON, Avro, Arrow, SQLite, Excel. No server, no import step, no extension to install before you can read a Parquet file. Same embedded model as SQLite and DuckDB, different defaults.

A few things we cared about while building it: It is one binary. Drop slothdb.exe somewhere, run it. It also runs in the browser. The WASM build is 1.3 MB and fits Workers 1 MB script cap in the edge variant.

It is fast enough to be worth the swap for analytical work. On a 5 query warm batch over 10M rows, SlothDB finishes in 138 ms. DuckDB 1.1.5 finishes the same batch on the same hardware in 540 ms.

It is also early and in development stages. v0.1.8. The Python wheel had a packaging bug that was only caught because a stranger filed an issue. So if you hit a rough edge, file one. SlothDB reads every one them.

Repo - https://github.com/SouravRoy-ETL/slothdb

heylo! open sourcing a plugin that I've been daily driving for a month. And I promise, I can't go back to vanilla Claude Code.

I'm a heavy Claude Code user, and a couple of things kept bugging me. Claude really likes to jump into code. And its responses are too long. I'd skim past walls of text just to find the recommendation. Plan mode helps, but it didn't go deep enough. The questions it asked were still implementation questions, when what I actually needed was alignment on the problem. So I made it stop.

Relay is a different output style and two Python hooks. When you describe what you want, it sits with you to fully understand your intent. Then it moves into the approach phase: walks through the options, pros and cons, and tells you which one it leans toward and why. It writes code only when you say so. When you're confident you and Claude are aligned, that's when you give the go ahead. And responses stay short. Feels like texting a staff engineer, not reading an essay. No more scrolling.

Relay helps you make the right decisions.

When you make a call together, it lands in .relay/sessions/decisions.jsonl in your repo. What you chose, what you rejected, why. Your next session picks up where this one ended, with the reasoning intact. Commit it, your teammate pulls, their Claude Code knows why you did it that way too.

The entire "brain" is one markdown file plus ~50 lines of Python. Read all of it in 5 minutes. MIT.

Install: /plugin marketplace add basegraphhq/relay-plugin then /plugin install relay@basegraph.

What I'd actually love feedback on: the preference learning bit. Relay is supposed to learn how you want to work via AskUserQuestion at real decision points, and silently from corrections you make. Tell me when it asks too much or too little.

Suprastructure is developing a deterministic instrument for verifying structural identity of complex hardware and logic designs.

Functional verification is necessary but no longer sufficient: it cannot detect silent architectural drift, supply-chain insertions, or hardware trojans that preserve observable behavior while altering the internal interaction graph.

Suprastructure generates partition certificates — mathematically verifiable fingerprints of a design’s interaction graph.

How it works: 1. Public seed.h + seed name ("instrumentality") + N=64 instantly builds the exact weighted directed graph W. 2. Our proprietary instrument solves for the canonical hierarchical partition tree T. 3. The result is published as a JSON proof containing T and its canonical SHA-256 hash.

Verification (anyone can do this in seconds): Rebuild W from the public GitHub seed (https://github.com/suprastructural/proof), walk the published tree to confirm every sup/sub split is admissible, then recompute the SHA-256. It must match the published hash.

Finding the correct tree by brute force would take >500 years. Verifying it takes seconds.

The free public portal demonstrates this with the synthetic seed above. Production engagements accept real design files (Verilog, SystemVerilog, AIGER, BTOR2, DIMACS, etc.) at any scale.

Suprastructure is an early-stage company with patents pending. We are seeking candid feedback from formal verification, hardware security, EDA, and high-assurance engineers.

If you have a few minutes, please try the synthetic seed and review the proof. I would greatly appreciate your input on clarity, usefulness for IP handoff / tapeout sign-off / supply-chain assurance, and what is missing.

Thank you for your time, consideration and feedback.

Why use Windows? Well, some people do. Why let Microsoft spy on your content? Why indeed? So, O&O Shut-Up (https://www.oo-software.com/en/shutup10) claims to cut the connection and telemetry between you and Microsoft. However, when Microsoft updates your computer, the settings are reset, and you have to run O&O ShutUp again.

Why not run OO which installs PowerShell 7+ if not installed, O&O Shutup, then it makes a scheduled tasks that runs the default settings after every Microsoft update?

Why futz around with re-running anything?

Just run OO once, and forget it.

Please file any bugs if you find them.

This isn't really going anywhere except to help internet users and my clients.

Hope it's helpful

PS - All of it generated with VS Code, Augmentcode AI (https://www.augmentcode.com/), OpenSpec (https://github.com/Fission-AI/OpenSpec), and beads (https://github.com/gastownhall/beads).