标签：#tor

When an AI agent runs a multi-step pipeline, every step creates temporary files. Article drafts, image uploads, JSON payloads, log files. Over fifty runs, these files accumulate. Some get cleaned up, some don't. And the ones that don't cause the next run to fail in confusing ways. I hit this exact problem with my publishing pipeline. A failed cleanup from run #12 left a stale devto_article.json in the working directory. Run #13 picked it up, parsed it, and published a draft with last week's title. The logs showed "JSON loaded successfully" — which was technically true. The file was valid JSON. It just belonged to the wrong run. The fix was a Guard class that sits between the pipeline and the filesystem. Every file the pipeline creates must be registered before the pipeline starts. Any file that appears without registration halts the pipeline immediately. Run identity gets embedded into every file, so even if a cleanup fails, the next run can tell the file doesn't belong. The Problem With Temp Files Temp files are invisible by design. You create them, use them, delete them. But when deletion fails — file lock, process crash, permission error — the file becomes a ghost. It exists on disk but nobody remembers it's there. The next run scans the directory, finds the ghost, and treats it as intentional. This is especially dangerous for JSON files because they're always valid. A stale manifest.json looks identical to a fresh one. The only difference is the content, and the loader doesn't check content provenance. Here's a concrete example from my pipeline: # The naive approach — just check if the file exists def load_manifest ( path ): if not path . exists (): return None return json . loads ( path . read_text ()) This code returns valid data from any run, any day, any context. It answers "can I read this file?" but not "should I read this file?" That distinction is the entire bug. The Guard Pattern The Guard class solves this by requiring every temp file to be registered

Angular is well served for icons. Material Symbols alone runs to thousands of glyphs, and the community wrappers pull in dozens of other sets on top. Geography is where the shelf runs out. You get globes and map pins, sometimes a set of flags, rarely the outline of Japan and almost never the six states of the GCC. GeoIcons ships country and area icons as Angular standalone components, 422 of them at the time of writing. Adding one takes three steps: install the package, import the component by its ISO code, and render its selector. npm i @geoicons/angular import { Component } from ' @angular/core ' ; import { Us } from ' @geoicons/angular/countries ' ; @ Component ({ selector : ' app-root ' , imports : [ Us ], template : `<geoicon-us aria-label="United States" />` , }) export class App {} See it live on geoicons.io → That is the whole path. Below: inputs, styling, accessibility, and picking an icon when you only know the country at runtime. Key takeaways Install @geoicons/angular , import each country by its ISO 3166 alpha-2 code in PascalCase, and add it to the component's imports array. The class is Us ; the selector you write in the template is <geoicon-us /> . Styling props are explicit inputs, because Angular has no rest spread. Everything else goes on the host element. Icons render as decorative unless you name them, so reach for aria-label only where no adjacent text says the country. Step 1: Install the Angular package Add the package to your project: npm i @geoicons/angular It needs @angular/core and @angular/common 15.1 or newer, plus rxjs 7 or newer. That 15.1 floor comes from hostDirectives , which the icons use to share their styling inputs and which landed in that release alongside standalone components . The package sets "sideEffects": false , so the CLI can drop what you never import. List three icons in a component and you ship three. Why that matters for icon libraries . Step 2: Import by ISO code Every country ships as a named export under its ISO

Coding agents usually do not drift because they are incapable. They drift because the task leaves too much room for interpretation. A request like “clean up authentication” sounds clear to a human who already knows the codebase. To an agent, it can mean anything from renaming one helper to replacing the entire authentication stack. The fix is not a longer prompt. It is a brief with four explicit parts : Outcome Context Guardrails Definition of Done Below is the exact structure I use. 1. State the outcome as an observable change Describe what should be different for the user or system when the work is complete. Weak: Fix the login bug. Better: When a user submits an expired magic link, show the existing “Link expired” message and offer a button that requests a new link without leaving the page. The better version gives the agent a destination. It does not prescribe the implementation, but it makes success testable. 2. Give only the context that changes the decision Context is useful when it removes ambiguity. It becomes noise when it is a tour of the whole repository. Useful context often includes: The relevant entry point or route The existing component or service that should be reused A similar implementation elsewhere in the codebase The command used to run the relevant tests A known constraint, such as backwards compatibility Example: The page is implemented in app/auth/verify/page.tsx . Reuse requestMagicLink() from lib/auth/client.ts . The existing error-message styles live in components/auth/AuthNotice.tsx . That is enough to start investigating without pretending we already know the final patch. 3. Add guardrails that define the change boundary Guardrails prevent a small task from becoming an accidental rewrite. A useful set might be: Do not change the public API. Do not add dependencies. Keep the current visual design. Do not edit generated files. Limit changes to the authentication flow and its tests. If a database migration appears necessary, stop and expl

If you've tried building a Chrome extension recently, you've probably hit Manifest V3 and spent an hour just figuring out why your background page stopped working. MV3 replaced background pages with service workers, changed how content scripts communicate, and made permissions stricter. The official docs are... not great. So here's the structure that actually works. The folder structure chrome-extension/ ├── manifest.json ├── popup/ │ ├── popup.html │ ├── popup.css │ └── popup.js ├── options/ │ ├── options.html │ └── options.js ├── content/ │ └── content.js ├── background/ │ └── service-worker.js ├── utils/ │ └── storage.js └── icons/ The manifest.json (MV3) The biggest MV3 gotcha: background scripts are now service workers. { "manifest_version": 3, "name": "Your Extension", "version": "1.0.0", "permissions": ["storage", "activeTab", "scripting"], "action": { "default_popup": "popup/popup.html" }, "background": { "service_worker": "background/service-worker.js" }, "content_scripts": [ { "matches": [""], "js": ["content/content.js"] } ] } Communicating between popup and content script This trips up almost everyone. The popup can't directly access the page DOM — it has to message the content script. // popup.js const [tab] = await chrome.tabs.query({ active: true, currentWindow: true }); await chrome.tabs.sendMessage(tab.id, { type: 'RUN_ACTION' }); // content.js chrome.runtime.onMessage.addListener((message, sender, sendResponse) => { if (message.type === 'RUN_ACTION') { // do something on the page sendResponse({ success: true }); } return true; // keeps the channel open for async response }); The return true at the end is critical — without it, async responses silently fail. Storage that syncs across devices Use chrome.storage.sync instead of localStorage. Here's a utility wrapper that makes it clean to use anywhere: const Storage = { async get(key) { return new Promise((resolve) => { chrome.storage.sync.get([key], (result) => resolve(result[key])); }); }, async set

Here is the whole method: get access to a real Linux machine, type commands on it for 30 to 60 minutes every day, and follow a plan that builds from navigating the filesystem up to running your own web server. Do that and you will be comfortable in four weeks and genuinely fluent in about eight. No experience required, no money required. The rest of this article is the specific plan: what to type each week, where to get a free machine you can safely break, what the three scariest errors mean, and how to tell you are actually improving. Why most people fail at Linux The pattern is nearly universal. Someone decides to learn Linux, finds a nine-hour video course, watches it at 1.5x speed, takes beautiful notes, and three weeks later cannot list the contents of a directory without checking those notes. Watching someone else type is not practice. It feels like learning because the explanation makes sense while you hear it. But command line skill is muscle memory wrapped around a mental model, and both are built one way: typing, failing, reading the error, trying again. An hour of reading about ls teaches you less than typing ls twenty times in twenty directories. Videos are fine as a preview. They are just not the workout. So flip the ratio: for every minute reading or watching, spend five with your hands on a keyboard. This article included. Read a section, then go type it. Two smaller failure modes show up almost as often. Trying to memorize everything Linux has thousands of commands. Working engineers lean hard on a core of about 25 and look up the rest without shame. The plan below teaches that core and nothing else. Fear of breaking things On a practice machine, breaking things is the goal, not the risk. A system you broke and fixed teaches more than ten flawless tutorials. Every option in the practice section makes the worst case "start over," which costs a minute. The four-week plan First, get a machine from the free options below (one minute to one afternoon, dep

How we added an on-demand translation help feature to our book translation platform, streaming LLM suggestions for tricky passages. At LectuLibre, our AI-powered book translation service allows users to upload EPUB or PDF files and get translations generated by large language models like Claude and DeepSeek. But we quickly noticed a pain point: automated translations, while fast, sometimes produced awkward or ambiguous results for culturally specific phrases, idioms, or technical jargon. Users wanted a way to get instant, contextual help for these tricky passages without leaving the platform. That’s when we set out to build the 翻译与转录求助 (Translation Assistance) feature — an interactive side panel where users can select any sentence or paragraph and receive alternative translations, explanations, and stylistic suggestions from an LLM in real time. In this article, I’ll walk you through the engineering challenge, the architecture we chose, and the specific code and trade-offs that made it work smoothly under production constraints. The Problem: Real-Time, Context-Aware Translation Help The core requirement was simple: a user highlights a piece of text in the translated book and clicks “Get Assistance”. Immediately, the system should stream back multiple translation options, a brief explanation of differences, and stylistic notes — all aware of the surrounding context, the author’s style, and the target language. Under the hood, this meant: Low latency : Users expect a response in under 2 seconds. Streaming : The LLM output can be long, so we needed to stream tokens as they are generated. Context awareness : We must include enough surrounding text from the book to ground the model’s response. No blocking : The main translation pipeline shouldn’t be affected; the assistance feature should exist as an independent async service. Cost efficiency : Avoid re-processing the entire book each time a user asks for help. Our Approach: Async FastAPI + SSE + Rate Limiting We run a P

Linear Regression: From Least Squares to Production-Ready Practice Tags : machinelearning , datascience , python , tutorial Linear regression is the first algorithm most people learn, and the one most people never study deeply. It is also the model you will still find in production after fancier algorithms fail, because it is fast, stable, and explainable. This article is not a "call .fit() and read the score" tutorial. We will cover the math, the statistical assumptions, the diagnostics, regularization, evaluation, production concerns, and the interview questions that separate beginners from engineers. Why Linear Regression Deserves a Second Look Linear regression is the foundation for understanding almost every other supervised model: Logistic regression is linear regression with a sigmoid on top. Ridge and Lasso are linear regression with constrained weights. Neural networks are stacked linear transformations with nonlinear activations. Tree models are judged against the same baseline: "can I beat a linear model?" More importantly, linear regression is still the right answer in many business problems. When you need to explain a prediction to a regulator, a client, or a finance team, a clean linear model with interpretable coefficients beats a black box. The Math: Least Squares and the Normal Equation Given features X and target y , a linear model assumes: y = X * beta + epsilon The goal is to minimize the residual sum of squares: L(beta) = ||y - X*beta||^2 Taking the derivative with respect to beta and setting it to zero gives the normal equation : beta = (X^T * X)^(-1) * X^T * y In practice, use the pseudoinverse ( pinv ) instead of the inverse, because X^T X may be singular or numerically unstable when features are collinear. import numpy as np def normal_equation ( X , y ): Xb = np . c_ [ np . ones ( X . shape [ 0 ]), X ] # add intercept beta = np . linalg . pinv ( Xb . T @ Xb ) @ Xb . T @ y return beta Three Equivalent Views of Least Squares 1. Geometric view

I was on a call last month with a startup CTO who had just gotten their AWS bill. They had built a beautiful RAG application: semantic search, conversational AI, the works. Their vector index was humming along with about 50 million embeddings. Then they hit product-market fit. Within six weeks, they scaled to 500 million vectors. Their monthly infrastructure costs went from $2,000 to $20,000. The real kicker? When we looked at the access patterns, over 80% of those vectors were queried less than once a week. They were paying hot-storage prices for data that was, by any honest measure, cold. The standard advice here is "just use a cheaper vector database." The more interesting question is: why are you storing all your vectors at the same temperature in the first place? The Cost-Recall-Latency Triangle Vector search forces a three-way tradeoff. You can optimize for cost, recall, and latency, but you only get to pick two. Want high recall and low latency? That costs money (in-memory HNSW graphs with full-precision vectors eating RAM). Want high recall at low cost? Latency goes up. Want cheap and fast? Recall suffers. Most teams pick a single point on this triangle and apply it uniformly to every vector in their index. That decision made sense when vector databases offered a single storage tier. It makes the same amount of sense as storing your entire filesystem on NVMe SSDs because some files need fast access. The conventional wisdom says you pick your point on the triangle and live with it. But the conventional wisdom was written before vector storage got interesting. The better approach: tier your vectors the same way you already tier your storage. Different access patterns deserve different economics. The same embedding that costs $0.12/month in RAM might cost $0.004/month on disk and $0.0002/month in object storage. When you have 500 million of them, those decimals matter. The Hot Tier: In-Memory HNSW and Exact k-NN For vectors that get hit constantly (your user-fa

How to Verify a SHA-256 Checksum on Windows, macOS, and Linux You download an ISO, installer, archive, or release binary. The publisher provides a long value such as: 9f86d081884c7d659a2feaa0c55ad015 a3bf4f1b2b0b822cd15d6c15b0f00a08 That value is a checksum, usually generated with SHA-256. Verifying it answers one practical question: Does the file you downloaded have exactly the same contents as the file the publisher hashed? A checksum mismatch can indicate a damaged download, an incomplete transfer, the wrong file version, or modified contents. Before verifying anything Get the expected checksum from a source you trust. Ideally, use the software publisher’s official website, release page, package repository, or signed checksum file. A matching checksum confirms that your file matches the data represented by the expected hash. It does not prove that the original publisher or website was trustworthy. If an attacker can replace both the download and the displayed checksum, they can make the two values match. For stronger authenticity verification, use a signed release when the publisher provides one. Verify SHA-256 on Windows Open PowerShell in the folder containing the downloaded file. Run: Get-FileHash ".\filename.iso" -Algorithm SHA256 Example: Get-FileHash ".\ubuntu.iso" -Algorithm SHA256 PowerShell returns something similar to: Algorithm : SHA256 Hash : 4A1F... Path : C:\Users\You\Downloads\ubuntu.iso Compare the value beside Hash with the checksum published by the download provider. Uppercase and lowercase letters do not matter in hexadecimal hashes. The characters themselves must otherwise match exactly. Compare automatically in PowerShell Instead of comparing two 64-character values manually, store the expected checksum and let PowerShell compare them: $expected = "PASTE_EXPECTED_SHA256_HERE" $actual = ( Get-FileHash ".\filename.iso" -Algorithm SHA256 ) . Hash if ( $actual -eq $expected ) { Write-Host "Checksum matches" } else { Write-Host "Checksum does not

I run a content pipeline that picks trending topics and publishes articles automatically. Last week I found out it had published the same story three times. Not the same title — the same exact topic, reworded each time. My dedup check was supposed to stop that. It didn't. Here's why, and how I killed the check. The Bug My pipeline had a similarity gate. Every candidate title got compared against the last 30 published titles, and anything scoring 0.58 or higher was rejected. Straightforward, right? from difflib import SequenceMatcher def jaccard_bigram ( a : str , b : str ) -> float : def bigrams ( s : str ) -> set [ str ]: return { s [ i : i + 2 ] for i in range ( len ( s ) - 1 )} x , y = bigrams ( a ), bigrams ( b ) return len ( x & y ) / len ( x | y ) if ( x | y ) else 1.0 def similarity ( a : str , b : str ) -> float : return max ( SequenceMatcher ( None , a , b ). ratio (), jaccard_bigram ( a , b )) THRESHOLD = 0.58 Here's the pair that slipped through. The candidate: 中国军队国际形象网宣片《当红》 And a title I had already published: 《当红》网宣片刷屏，普通人看到的中国军人是什么样 Same film. Same topic. Third time it was being covered. Watch what the algorithm did: candidate = " 中国军队国际形象网宣片《当红》 " published = " 《当红》网宣片刷屏，普通人看到的中国军人是什么样 " print ( similarity ( candidate , published )) # SequenceMatcher: 0.187 # jaccard bigram: 0.185 # max: 0.187 < 0.58 -> PASSED 0.187. The gate let it through with a five-fold margin to spare. Why It Failed The name 当红 is the same in both titles. That is the whole topic. But the algorithm does not care about that. SequenceMatcher matches in order. In the published title, 当红 sits at position zero. In the candidate, it is at the end. Reordered tokens break the match, so the ratio collapses to the shared fragments — 网宣片 plus the generic words around it. The bigram fallback does not save you either. Jaccard over character bigrams measures surface overlap, not meaning. Five shared bigrams out of twenty-seven total. 0.185. It "proves" the titles are unrelated because most of

If your script suddenly started printing this: >>> import yfinance as yf >>> df = yf . download ( " NG=F " , period = " 1mo " ) 1 Failed download : [ ' NG=F ' ]: YFPricesMissingError ( ' possibly delisted; no price data found ' ) …you didn't break anything. NG=F (the natural gas futures ticker on Yahoo Finance) periodically stops returning data for everyone, and futures tickers get hit harder than stocks. This post covers why it happens and the three fixes that actually work, ordered from "quick patch" to "never deal with this again." 1. What the error actually means yfinance is not an official API . It's a (great) community library that scrapes Yahoo Finance's internal endpoints — the same ones Yahoo's own website uses. Yahoo doesn't document them, doesn't promise they'll keep working, and changes them whenever it suits their frontend. When Yahoo changes something — an endpoint, a rate limit, a response format — yfinance breaks until its maintainers reverse-engineer the change. Futures symbols like NG=F and GC=F are the most fragile: they've had recurring gaps and failures reported over the years, for example #2620 (missing recent data for NG=F/GC=F) , #2635 (whole missing days in futures history) and the evergreen #865 "Futures only work sometimes" . So: "possibly delisted" almost never means delisted. It means "the scrape came back empty." 2. Fix #1 — the quick patches (works today, breaks tomorrow) Three things fix most transient failures: Upgrade first. The maintainers usually patch Yahoo changes within days: pip install -U yfinance Retry with backoff. Failures are often intermittent rate-limiting, not hard breaks: import time import yfinance as yf def download_with_retry ( ticker , retries = 3 , wait = 5 , ** kwargs ): for attempt in range ( 1 , retries + 1 ): df = yf . download ( ticker , progress = False , ** kwargs ) if not df . empty : return df print ( f " attempt { attempt } came back empty, retrying in { wait } s… " ) time . sleep ( wait * attempt ) rai

Liquid syntax error: Unknown tag 'endraw'

JDK 24 removed the monitor-related carrier-thread pinning that stalled Netflix and similar teams on Java 21. What has replaced it on JDK 25 LTS is downstream-resource saturation: The bottleneck moved and now demands explicit bounding in application code. This article maps the failure modes that surface after virtual-thread adoption and gives a practical sequence backed by a public benchmark. By Sandeep Bharadwaj

Article Summary Bulk-generating product pages with AI looks simple: send product attributes to a model and ask it to write persuasive copy. In practice, this approach often creates invented claims, mismatched specifications, repetitive content, prohibited wording, and formats that cannot be published across different sales channels. A production-ready system is not a loop that repeats one prompt. It is a content pipeline that combines product-data cleaning, factual constraints, structured generation, rule-based validation, human review, and multi-channel publishing. This guide provides a practical data model, prompt template, JSON output schema, Python batch-processing example, and quality-control checklist. Why Direct AI Product-Copy Generation Often Fails A common workflow is to copy a product name and a few attributes from a spreadsheet, then ask: Write an attractive product detail page. The model may produce fluent text, but fluent text is not necessarily accurate product content. Five problems appear repeatedly. The source data is incomplete Many product spreadsheets contain only: SKU; product name; price; one or two specifications. A useful product page may also require target users, use cases, materials, dimensions, packaging, warnings, warranty terms, and verified benefits. When these facts are absent, a language model may fill the gaps with plausible but unsupported details. Facts and marketing claims are mixed together “Made with 304 stainless steel” is a factual attribute. “Designed for everyday durability” is a restrained interpretation. “The safest and most durable cup on the market” is an unverified claim. If the system does not distinguish facts from acceptable marketing language, the model may present assumptions as product truth. Every channel has different requirements The same product may need: an SEO title and meta description for a direct-to-consumer website; marketplace-style feature sections; Amazon bullet points; a short video script; social-

When building modern Python applications—whether scraping web pages, fetching data from external APIs, or querying databases—IO-bound operations often slow down execution. Python’s concurrent.futures module provides a high-level, elegant interface for running tasks asynchronously. In this guide, we'll break down what Futures are, why you need them, and how to use them effectively using a practical e-commerce product service. What is a Future? A Future represents an eventual result of an asynchronous operation. When you launch an expensive, long-running task concurrently, your program doesn't pause to wait for the output. Instead, it instantly gets back a Future object —a low-cost proxy or standard "claim ticket." The Future acts as a placeholder for a result that hasn't been computed yet. It keeps track of the task's execution state ( PENDING , RUNNING , CANCELLED , or FINISHED ). Once the task finishes, the Future stores the return value or any exception thrown during execution. Why are Futures Needed? In standard synchronous Python execution, calling a function blocks your main thread until that function finishes: Task 1 (2s) ──> Task 2 (3s) ──> Task 3 (1s) = 6 seconds total When dealing with IO-bound operations (like waiting for network responses or reading disks), your CPU sits completely idle during those delays. By offloading tasks into background threads or processes via Futures, your application can run multiple IO operations simultaneously: Task 1 (2s) [████████] Task 2 (3s) [████████████] Task 3 (1s) [████] ----------------------------------------- Total Time: 3 seconds (time of longest task) When Should You Use Futures? IO-Bound Workloads: Scraping multiple web pages, batch-calling microservices, querying multiple databases, or fetching images concurrently ( ThreadPoolExecutor ). CPU-Bound Parallelism: Performing heavy mathematical operations or image processing across multiple CPU cores ( ProcessPoolExecutor ). Decoupled Workflows: When you want to trigg

I recently discovered you can build a fully working Flexmonster pivot table right inside a Claude...

Apple said a slowdown in mobile gaming and changes to the App Store’s business model — including court-ordered payment rule changes in the U.S. — weighed on its services business, even as the company topped 1.5 billion paid subscriptions.

You are a software engineer. Your craft honed through years of careful practice. Then suddenly, there are these chatbots and agents. Overnight, your colleagues got a new title on LinkedIn: "AI engineer". Some are already SENIOR AI engineers. You're curious about this new world, and might want to catch up and become part of it yourself. If this is you, then join me on this tour through the concepts and patterns that make up the field of AI engineering. We will find that AI application development is mostly 'just' software engineering, applied to one genuinely strange new non-deterministic component: the LLM. During the tour, we build a real application, end to end. Every article adds a new layer. We link the new patterns and words to existing software engineering concepts you already know. Before take-off, I'd like to establish one vocabulary rule used throughout: "the model" means the LLM itself (large language model, like GPT or Claude), and what AI engineers build around it will be referred to as "the application", "the agent" or "the harness". What we're building As I work at a payments company myself, I figured I'd stick to my domain. PayIQ, the application we build, is an assistant for merchants to perform payment operations: issue refunds, defend chargebacks, calculate processing fees. Give it a charge amount and a payment method, and it computes what a refund actually costs (spoiler: more than the refund amount). Ask it whether a chargeback is worth fighting, and it does the expected-value math using your knowledge base. Ask it something it can't responsibly answer, and it asks for what's missing. No guessing, no hallucinations. By the end, PayIQ will have structured outputs that can be consumed by other systems, a tool belt of financial calculators, retrieval over a knowledge base, an agent loop with persistent memory, an orchestration graph with steps the model cannot skip, token streaming behind a FastAPI service, a regression eval suite, and layered injec

If you just want the recommendation: for the ordinary AI-generated image an inference job hands back — a 2 to 8 MB PNG — do one plain object PUT into your S3-compatible storage and stop there, because multipart upload only earns its complexity when a single artifact is big enough that losing a transfer halfway through costs you real money to redo, which for my team starts somewhere north of 100 MB. Everything below is about that threshold, and about the operations bill you pick up the moment you cross it. I run the platform roadmap for a team that renders a few hundred thousand images a month, and I count pages before I count features, so read the rest with that bias in mind. Should I use multipart upload for large AI-generated images, or a single object PUT? Multipart solves two narrow problems: a payload too awkward for one HTTP round trip, and a transfer you refuse to restart from byte zero. A 6 MB PNG has neither problem. The shape of the flow is always the same wherever you run it. You start a multipart upload and get back an upload id, you push each part under that id, you collect the returned ETag and part number for every one of them, and you send the finished list back in a complete call that stitches the object together server-side. Parts have to be at least 5 MiB on Amazon S3 and on every S3-compatible store I've tested against, with the final part exempt, which already tells you the feature was designed for objects measured in hundreds of megabytes rather than for a batch of thumbnails. Where it genuinely pays off in an image pipeline is the long tail: a 4-gigapixel tiled upscale, a nightly ZIP export of a customer's whole render history, a raw latent archive somebody in research wants kept for a year. Those are the jobs where a dropped connection at 80% is a real incident and not a shrug. For everything else, one put is one line of code and one thing to monitor. There's a second cost that people underrate, and it's the one I'd argue about in a design re

Whether you are using off-the-shelf AI coding tools like Claude Code and Cursor or building custom autonomous AI agents with TypeScript and LLM APIs, you hit the exact same fundamental wall: AI agent amnesia . As an agent user , you spend forty-five minutes explaining your architecture, deployment quirks, and database rules. The agent writes brilliant code. You close the CLI or tab, open a new session the next morning, and the agent suggests the exact legacy library you rejected yesterday. As an agent builder , you struggle to keep your custom agentic loops focused. As multi-step agent trajectories expand, LLM token limits force context compaction, wiping out subtle rules and past decisions while escalating API costs. The intelligence is real. The amnesia is structural. Context Windows Are Working Memory, Not Long-Term Memory The AI industry’s standard reflex to agent amnesia has been pushing context windows to 1M+ tokens. But a context window is working memory (RAM), not long-term storage (disk). Relying on massive context windows introduces three critical engineering bottlenecks for both users and builders: Context Compaction Destroys Rationale : When a session reaches token limits, agents automatically compact their context history. Compaction summarizes conversations into short summaries, quietly wiping out subtle architectural constraints, edge cases, and past decisions. Context Drift & Attention Loss : LLMs struggle with needle-in-a-haystack attention degradation when context windows are stuffed with 100k+ lines of raw conversation history. Escalating API Costs & Latency : Re-sending full project transcripts on every prompt burns tokens rapidly and adds seconds of input processing delay for users while skyrocketing LLM bills for agent builders. Agents do not need larger transcripts. They need a durable, inspectable, versioned memory layer . Why Vector Databases Fall Short for Local & Workspace Agent Workflows When developers and AI engineers realize raw contex