Context window limits & the model forgetting earlier context

At some point, keeping track of the agents became a problem of its own. I kept forgetting which agent was working on which branch/worktree, which sessions were waiting for me, and which changes I still needed to review among the many chats I have open.

lcx · Nov 2, 2026 · comment in “Repobs – A local dashboard to manage agent worktrees and branches”

Prior to that, I've built an ai workflow engine and a platform for both minimizing hallucinations and working with documents that are far too large for existing context window inputs or outputs using a novel, patented process that I invented; and applied these tools to various contracted projects requiring analysis and and cross-checking of information and other forms of integrity across textual and video/audio corp…

poeticsilence · Nov 2, 2026 · comment in “Ask HN: Freelancer? Seeking freelancer? (November 2026)”

Prior to that, I've built an ai workflow engine and a platform for both minimizing hallucinations and working with documents that are far too large for existing context window inputs or outputs using a novel, patented process that I invented; and applied these tools to various contracted projects requiring analysis and and cross-checking of information and other forms of integrity across textual and video/audio corp…

poeticsilence · Nov 2, 2026 · comment in “Ask HN: Who wants to be hired? (November 2026)”

An issue with a VSC extension. And the path for this issue to even be a thing possibly a weakness in the way the extensions marketplace works (so frustrating that marketplaces and app stores are like spam folders in that you have loads of entries attempting to convince you that they are something they aren't.) And I often forget that Cursor even has an IDE, and that the IDE is their own release of VSC.

gexla · Oct 31, 2026 · comment in “Backdoor in Claude Code for VS Code extension on the Cursor Marketplace”

Three years ago AI generated code could basically not be used at all, and I had to rewrite just about every line. Today I have to watch out for generated code making steps in the wrong strategic and architectural direction, and over engineering, no longer hallucinating nonexisting APIs and forgetting to check the auth token.

thomashabets2 · Oct 23, 2026 · comment in “Is this the new cyber normal?”

An example: a year back I wanted to show that AI doesn't hallucinate that much. Here's like a 20 comment long conversation with a person who believed that ChatGPT hallucinates easily.

simianwords · Oct 19, 2026 · comment in “It's still ok to enjoy programming”

Fair enough yeah, I think the thing that is annoying is that the service design is explicitly disincentivizing keeping an eye on the tasks, with all of the "set it and forget it" features.

bulder · Oct 18, 2026 · comment in “DeepSeek search and fetch – $0.75 per 1000 successful requests”

With respect, the wins you're describing are exploits that the next iteration of bots will fix. The "AI forget the Reaver exists", "AI doesn't understand how dark swarm works", "AI doesn't know what nukes are" are bugs that can be corrected, not big gaps in reasoning.

hitekker · Oct 18, 2026 · comment in “Show HN: Gemini 4 Argon plays StarCraft”

Describing the opinions of large groups of people isn't clear-cut enough to justify calling it fallacious. I personally think that most anti-AI people I've met struggle to articulate good reasons to hate AI (here good = internally consistent) and I do think they've lost the plot.

SR2Z · Oct 18, 2026 · comment in “I'm Fucking Done with Slop”

Also single seat workspace user. Half the time my Gemini forgets I connected Gmail to it and hallucinates false instructions to enable it.

radicality · Oct 13, 2026 · comment in “Gemini 4 Argon”

And then time passes and the issue is either mitigated or circumvented entirely. Hallucinations are one, "AI can't scale anymore" is another (remember that one?), limited context windows, strawberry problems, glue on pizza, sycophancy.

glenstein · Oct 12, 2026 · comment in “GLM-Legal”

It’s already the case that when I write code by hand I forget to do stuff that an AI with good instructions would never forget. I struggle with Boolean conditions that get more complex than a few operators while the AI can easily explain and track things like that.

brabel · Oct 9, 2026 · comment in “Yes, and”

All the LLM issues I face are related to incorrect priorities, choosing the wrong tradeoffs, UX and UI flaws or incorrect assumption about requirements. Why would I want to fill up my context window with useless type annotations and waste tokens generating them?

jongjong · Oct 6, 2026 · comment in “Friendship ended with Deno, now Node is my best friend”

Trying to optimize tools in the age of AI is like putting lipstick on a pig. - It's more verbose than most other languages; waste of tokens, context window, time and AI reasoning capacity.

jongjong · Oct 1, 2026 · comment in “How to speed up the Rust compiler in September 2026”

I canceled Claude for my personal account due to the Astra hype, but I'm back after a single month and the absolutely ridiculous context window is one of the many reasons. I barely compact at work in a very complex monorepo (neither with Fable 5.1 nor Opus 5.5), and yet in my personal greenfield project Astra keeps compacting all the time, to the point of it being unusable.

kaoD · Sep 30, 2026 · comment in “GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price”

It has forced me to use persistent project documentation for everything. Maybe that's not a bad thing but unless it reads all the documentation after every compaction (and uses half its cache), it goes off the rails.

Gareth321 · Sep 30, 2026 · comment in “GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price”

Its really not tiny; you can't compare Claude to GPT, they have honestly diverged enough that as the other reply said, 256k GPT is about equal to 1M Claude. The compaction is slightly annoying, and you can turn it up to 1M as you said if you truly need everything in context, but otherwise it's perfectly serviceable

rrvsh · Sep 30, 2026 · comment in “GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price”

- it said it would ask me to approve/amend the suggested PR message, it never did and fired off right away - it keeps forgetting the changes it did itself; no context compaction was used

cromka · Sep 28, 2026 · comment in “Sonnet 5.5”

YMMV of course but I was shocked at how quickly I burned through Codex usage. On the context window, I feel so cramped on Codex, compacting happening every time I turn around is annoying.

joshstrange · Sep 22, 2026 · comment in “GPT-6 Sol and Luna”

Don't forget that Opus 5 was tracking fable on many benchmarks, yet it was borderline unusable for any coding work.

glub · Sep 22, 2026 · comment in “Claude Opus 5.5”

Realistically, programming domains fall into two major categories: "ship it and forget it" (one-offs) and continuous services. I agree with the OP's point that AI struggles to understand boundary delineations.

jdw64 · Sep 22, 2026 · comment in “AI Has No Wisdom and Neither Will You”

Most memory system solutions focus primarily on the storage and retrieval aspects (vector search/RAG/graphs/ Markdown files, etc.). After months of storing memories (coding 8+ hours a day produces a lot of memories) these might hallucinate your reasoning model and they can clutter your context window.

mrsalty · Sep 14, 2026 · story in “Show HN: Slowave – local adaptive memory for coding agents”

They do not perform benchmarks to evaluate whether new "features" are any good and instead bloat the context with more and more tools that are rarely useful and often confuse models. Eventually, the context window got so bloated that they resorted to hiding function bodies in large files, which is of course a stupid idea because then the LLMs have to use other tools to read the files, wasting even more tokens, or ha…

gpugreg · Sep 14, 2026 · comment in “Notes on gotchas while migrating 35kb preprompts from Opus to self-hosted Ollama”

3. Saying AGENTS.md can enforce anything at all is a pipe dream, I am struggling with this all the time, in a long session the tokens from AGENTS.md are so far in the back of the context and so diluted that LLM's "attention" doesn't register them.

aytigra · Sep 14, 2026 · comment in “A computational constitution to stop LLM agents from bricking servers”

The model is asked to summarize a massive blob of mixed-relevance information when its reasoning quality is already at its lowest. This compaction leads to forgetting critical constraints, makes way for hallucinations of past decisions, and introduces noise that makes the new context unreliable for long-horizon work.

shaurya-sethi · Sep 10, 2026 · story in “Show HN: Nightshift – Rust CLI to Orchestrate GitHub Issue Resolution with Dags”

- Agent unable to keep making updates without looping forever (the complexity of the code exceeds the agent's context length). Human doesn't understand either so can't fix.

odo1242 · Sep 2, 2026 · comment in “AI Agents and the Refactoring That Never Happens”

There is an in between here that developers seem to keep forgetting: Design the software architecture yourself — do some whiteboarding, figure out your abstractions and your UML, then write that in your prompt for the LLM to implement through a plan.

bitwizeshift · Aug 31, 2026 · comment in “AI-written code is still your code”

I'd like to work with a corpus offline (internal university research data) and I'm hoping I can get everything done without the data leaving the premises. I guess the biggest bottleneck is going to be for the context window size which won't be able to fit too many result "hits."

ivansavz · Aug 26, 2026 · comment in “RAG Is Simpler Than You Think”

The only problem I have know is how pi compaction works, which is forcing full prefill which takes time and it is erroring a lot. I wrote my own compaction that should remove full prefil but it does not work.

npodbielski · Aug 25, 2026 · comment in “Anthropic's best AI model struggles to attract users as cheaper tools thrive”

.. except that AI is a retarded intern, constantly forgetting stuff unless reminded all the time. That's the most frustrating part, really.

theamk · Aug 23, 2026 · comment in “Treat AI Like an Intern, Not Software: A Stanford Professor's Guide”

Using AI agents is amazing, but in long-term, I started to have more and more problems about the context of them. They keep forgetting what we are working on, opening a new chat session clears all their context, my CLAUDE.md had like 1000 lines.

dat999zx · Aug 22, 2026 · story in “Show HN: Knowl – CLAUDE.md hit 1000 lines, so I built memory that prunes itself”

Oh yeah, I'm still amazed how good the current iteration of models are for coding (I have a fear it's too good to be true - so will get taken away..). Exactly a year ago I switched from GPT 5 to Gemini just because the coding with R language was terrible; and even with Python it kept forgetting and mixing basic stuff.

5555watch · Aug 19, 2026 · comment in “Cerebras CS-4”

I mean just look at the benchmarks, even though Gemini 3.7 Flash performs well on the DeepSWE 1.1, Luna (Max) still performs way better. I personally have stuck to Luna (Xhigh) because its been more than enough and does not bloat up the context window too fast with reasoning tokens.

Alifatisk · Aug 13, 2026 · comment in “Gemini 3.7 Flash”

The other day, Sol called Opus 5's handoff (a skill I have that is basically a compaction, but just written to a file not tied to one LLM) "incoherent", that was a new one. Opus 4.8 or Fable (at great expense) are the only ones that aren't frustrating for me.

pmarreck · Aug 12, 2026 · comment in “Grok 4.6”

I think the places most afraid from cultural dilution are the ones that needed it the most. It's also so funny how folks in the US keep forgetting that they literally used to be immigrants of all sorts of origins and what made them so successful is leveraging the diverse perspectives.

kwdev · Aug 8, 2026 · comment in “New Amazon Data Center Is Set to Have the Most Polluting Power Plant in the U.S.”

Why should I even talk to it?? It does not follow anything I say, it does not read it's rules, it doesn't read the code in the amount needed, after one prompt it again forgets what I wanted - it's, sorry, shit.

theGeatZhopa · Aug 5, 2026 · comment in “Opus 5 is lazy (A Twitter thread)”

Obviously you shouldn't forget about traditional tooling for formatting and linting, as well as static code analysis and having test coverage that approaches 100%. It might be annoying to do manually, but AI has no issues with refactoring code to make it more testable and eventually will catch some issues that way.

KronisLV · Aug 1, 2026 · comment in “AI doesn't generate working products, that's still your job”

Make subagent invocations into tool calls to that agent. Externalize the tool calls themselves to CLI utils, maybe block native tools completely (unrelated but relevant, I just added the AskUserQuestion tool to the global deny list couple days ago because Claude occasionally forgot to honor my standing order to prefer plain text, and that tool is extremely annoying as it creates a gap in the dialogue) and use 3p alt…

skeledrew · Jul 31, 2026 · comment in “The session you cannot take with you”

You will find that a lot of problems suddenly go away: long context degradation vanishes (which Qwen 3.6 will still do inside it's 250k context, especially when both model and KV are below Q8), forgetting what you said or it said earlier goes away, losing the plot half way through goes away, overreliance on cliches (Qwen has its own form of Claudisms, but are more subtle and less grating) also goes away, hallucinati…

DiabloD3 · Jul 29, 2026 · comment in “Handbook.md shows that long policy documents do not reliably govern agents”

Any new to share on file upload support? We have shipped a MCP server and it has been really frustrating to observe MCP clients fumbling around with base64-encodings, polluting their context window with binary data.

jakobgm · Jul 28, 2026 · comment in “MCP 2026-07-28 Specification: transport going stateless”

react based shit lost the plot so hard. i've used it quite a bit, with other beefy but impressive js libs like aggrid and stuff. i'm just tired of client-browser overwork in general that feels like such a mess. having so much of a webapps logic live in the front end feels completely preposterous to me these days. we've only ever needed front end to just.. be the front end. look nice, provide the interface, give peop…

trueno · Jul 27, 2026 · comment in “Htmx 4.0, the first JavaScript library to release exclusively on the Game Boy”

Also, don't forget the cookie basically also dug a hole in the security setting. So it's really a balance between avoid annoying the user and protect your site.

nirui · Jul 26, 2026 · comment in “Cloudflare's new AI traffic options for customers”

It’s funny, I’d genuinely say that Gravatar improved for quite some time. In the old days it was fine, then for quite some time under Automattic when they started fiddling with it it was ridiculously bad and for a while even literally unusable due to broken tie-in to WordPress.com accounts or something (I forget), then they finally got it together again so that it was arguably better than before they took it over, w…

chrismorgan · Jul 25, 2026 · comment in “Re: Bye Bye Gravatar”

People keep forgetting that over the last 6+ months a lot of increased action has been taken by OpenAI and Anthropic to detect and combat distillation.

benjiro29 · Jul 22, 2026 · comment in ““We have information that Moonshot distilled Fable for the development of K3””

You can run ’codex debug models' into jq! I forget whether it's documented or not, and it is kind of annoying that that's how you find it, but it does tell you.

tekacs · Jul 19, 2026 · comment in “OpenAI reduces Codex Model Context Size from 372k to 272k”

They proudly claim that every AI project they've observed over the past year and a half had a 0% success rate, and that they've rejected all AI implementation work. While this is evidence that the market is crazy, at its core, it's a painful confession that they have no engineering expertise to implement and control modern AI architectures like RAG, Agentic Workflow, and context window optimization to meet business …

jdw64 · Jul 19, 2026 · comment in “AI Mania Is Eviscerating Global Decision-Making”

And god forbid I don't, the amount of time I've lost to Claude just sending it and then lying is not worth it when 5.5 will just do stuff, compact 2-3 times in the task and still not go insane. This whole 1 million context window is a lie after 300k Claude degrades to unusable, and compact doesnt help.

what-the-grump · Jul 18, 2026 · comment in “Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal help?”

I cannot imagine how much context window it needs for this. I am struggling with local AI in Pi to even get small (500-1000 lines of code) programs written without overrunning a 220k context.

Schlagbohrer · Jul 17, 2026 · comment in “Kimi K3: Open Frontier Intelligence”

I actually think Claude gets smarter the longer a conversation goes on (up until compaction). I have noticed Claude trying to wrap up long sessions, and it's extremely annoying.

Wowfunhappy · Jul 14, 2026 · comment in “Claude Code plugin that plays a Mr. Meeseeks voice line whene Claude is waiting”

The challenge I think for codex is the restriction on context size and the constantly rolling compactions. They are less aggressive or disruptive but it is still annoying you can’t force a 1mm context window on a 1mm context window model.

fnordpiglet · Jul 12, 2026 · comment in “Claude Code May–July 2026 weekly limits promotion”

I don't think it has almost anything to do with vibe-coding. When I tried Haskell and OCaml some 3 years ago, they were infuriating in terms of tooling and just needless friction (don't ask, forgot, but it was mostly around package management and just entering a project and be able to run the program was a saga).

pdimitar · Jul 12, 2026 · comment in “After 7 years in production, Scarf has reluctantly moved away from Haskell”

Waiting for the VM to get set up when responding to a task made them feel like they were only for fire and forget workflows, and I missed going back and forth with the agent and using plan mode. On top of that, the whole VM was a black box and I had no way of verifying the agent's work besides checking out its branch and testing its changes myself, which got annoying super quickly.

pranav100000 · Jul 11, 2026 · story in “Show HN: Aether – Run Claude Code, Codex, or OpenCode in devboxes you can watch”

As for context, in my experience Claude starts trying either to do maximum work with minimum tokens when it's approaching limit, or it starts deferring useful work while doing busy work. Both result in a mess and complete loss of traction after compaction.

0x457 · Jul 10, 2026 · comment in “GPT-5.6 Sol Ultra produces proof of the Cycle Double Cover Conjecture [pdf]”

If I forget about a conversation it's been cognitively difficult for me to track down which tmux tab = the correct remote control session in the Claude GUI. Remote control does come in super handy for pasting images in when some shell environments where pasting doesn't work for...reasons.

xbryanx · Jul 8, 2026 · comment in “Herdr: One terminal to rule them all”

RAG, as far as I understand, is a term that came about before LLM tool-calling was as prevalent. Your options were to have an LLM hallucinate up a response, or instead do a [document -> chunk -> embedding -> vector db -> query -> context window] pipeline.

teaearlgraycold · Jul 7, 2026 · comment in “Pruning RAG context down to what the answer actually needs”

Thankfully the appeal form worked, but by then I was using a different Claude account praying they didn’t ban me again. There’s an uncensored model floating around that you can run locally with llama.cpp: https://www.reddit.com/r/LocalLLaMA/comments/1rq7jtm/qwen3535ba3b_uncensored_aggressive_gguf_release/ it’s annoying to use since you run out of context window quickly, and it’s certainly not able to be deployed in …

sillysaurusx · Jul 7, 2026 · comment in “Fable turned reMarkable into Tom Riddle's diary from Harry Potter”

Same. As someone who is OK with a small amount of maintenance every N months, but keeps forgetting how things are setup or what I did, moving absolutely everything into Nix and running NixOS made things a hell of a lot simpler when you come back after 6 months and can easily find where and what to change, as long as you take care to declaratively set things up via Nix as much as possible, and use git.

embedding-shape · Jul 1, 2026 · comment in “I don't maintain my homelab, it maintains itself”

I don’t understand how this is a different result than giving any LLM a task that is not completely grounded? I’ve observed this in coding tasks, if I forget to include a file referred to in the spec, the LLM will just hallucinate a version of it and my results suck.

seanmcdirmid · Jun 28, 2026 · comment in “I used Claude Code to get a second opinion on my MRI”

I've started building Lore because none of the existing "memory" solutions solved my real problem: my agent stopped every 5-10 minutes to "compact" and forget every important thing we covered before these compaction stages.

BYK · Jun 26, 2026 · comment in “Why memory is not enough: you need context management”

Aside from the obvious AI slop writing, nothing in the example he showed guarantees that this is "improving itself". Each iteration of the answer, as given back by the LLM can get increasingly worse, especially as the context window grows, or as something that shouldn't be there (e.g. a hallucination) in the previous answers becomes part of the later answers' focus, or the "roll of the dice" that's an LLM answer get…

coldtea · Jun 24, 2026 · comment in “The loop is here (reply to "The Coming Loop" by mitsuhiko)”