🔥 Trending on HN

How Can a Computer Helper Remember Each Project?

3 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
AI agents(AI agents)

Software helpers that carry out tasks for a user.

context(context)

Past information and rules needed for a task.

Hacker News(Hacker News)

A website where people discuss technology.

What happened

A user posted an Ask HN question about carrying work history across projects. The writer uses AI agents, software helpers that carry out tasks for a person, across several clients. A client is a customer or organization. The writer says the biggest time cost is explaining the same context again.

Here, context means the past information and rules needed for a task. The post names material from chats, email, and tickets. It also names earlier decisions and each client’s rules. The user asks how people bring those pieces into a new session.

The thread drew 98 points and 58 comments in the supplied snapshot. Those numbers show community attention. They do not prove that the question or any proposed solution is correct. The Hacker News post is the source for both the question and that discussion listing.

Background

The post is not a product launch. It is a request for working methods. The author asks what people give an agent when a new session begins. They also ask whether people use more than one kind of agent. If so, they want to know how people move context between agents.

The author asks about client boundaries too. Information for one client may follow different rules from information for another. The post asks how workers keep those settings separate. It also asks what happened when an agent forgot a decision or used an approach that had already been rejected.

Finally, the author asks whether people bought or built a tool. The author says they are exploring whether to build something themselves. That places the thread at the question-gathering stage, not at the proven-solution stage.

Why it matters

The post treats continuity as a work problem. A useful answer may depend on earlier choices, not just on the current request. If a new session lacks those choices, a person may need to repeat the background. If an agent carries the wrong project history, it may follow the wrong rules. Those are risks raised by the question, not confirmed incidents in this source.

This makes context management more specific than simply asking an AI to remember more. Someone must decide which past details belong to which project. They must also decide what another agent needs during a handoff. The post asks people to explain those choices in real work.

What we can confirm

We can confirm that one user works with agents across several clients. We can confirm that repeated explanations are a major concern for that user. We can confirm five areas of interest: starting a session, switching agents, separating clients, handling forgotten decisions, and choosing tools.

We cannot confirm a winning workflow from this post. It reports no benchmark, product result, or measured time saving. Hacker News activity shows that the topic attracted attention. It does not turn one user’s question into a survey of all AI users.

What remains unknown

The source does not establish which records work best. It does not show how much time a method saves. It does not measure how often agents forget decisions. It also does not show how users verify that one client’s information stayed separate.

It is also unknown whether the author will build the tool they are considering. A future report would need to show what gets stored, how agents receive it, and how mistakes are detected.

What to watch next

The next useful development would be a concrete workflow or tool from the author. It could show how notes move from a chat, email, or ticket into a project record. It could also show how another agent receives only the rules it needs.

Until then, the news is the question itself. The thread presents continuity as a practical limit on multi-project AI work. It asks whether better handoffs can reduce repeated explanations without mixing client contexts. The attention on Hacker News shows interest in that problem, not a final answer.

Source: Hacker News post

💬 Managing AI-Agent Context with Documents, Retrieval, Isolation, and Checks

The dominant pattern in the thread is not one enormous shared prompt, but project- and client-specific knowledge stored in documents and retrieved only when needed. The performance figures and failure reports below are commenters’ own claims, not independently verified results.

  • Many commenters organize Markdown, Obsidian, wikis, ADRs, issue trackers, and design documents by project or client. Small AGENTS.md or CLAUDE.md files act as entry points or indexes to the detailed material.
  • To carry work across sessions, some maintain a topic.md or handoff note at the end of each session, move older detail into an archive, and keep the active context compact.
  • One workflow combines an LLM Wiki with QMD semantic search and searches the knowledge base before significant work. A gatekeeper.py script checks record keeping, wiki updates, documentation, and linting before allowing a round to count as complete. Its author says this costs tokens but makes project stages reproducible.
  • Zaterka reports that, on the same benchmark, LLM Wiki increased token use by 23%, graphify increased it by 5%, and a hybrid tool called wayfinder reduced it by 30%. These are self-reported comparisons, not independent measurements.
  • Another commenter reports that cv’s pruning and compaction saved 40% of tokens in one evaluation. That result is also specific to one reported environment.
  • To prevent client context from mixing, commenters use separate accounts, spaces, directories, repositories, or agents. One design layers context as root, agent, client, and project, exposing only the relevant branch.
  • The author of Varda reports a failure that cost a container holding a sample database, and describes using agent-selection controls, orchestration, and sandboxing to limit the blast radius.
  • A strong counterargument is that well-organized Markdown, folders, and a master index may be enough. An index can reduce the amount of scanning needed in a cold session, so a dedicated service is optional rather than necessary.

mature digest at 58 comments (revision 1). We fetched 60 comments and sampled 60 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

How Can AI Helpers Keep Each Project Straight?

📰 Full story: How Can a Computer Helper Remember Each Project?

A Hacker News post asks how computer helpers carry old decisions between projects.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
AI agents(AI agents)

Computer helpers that can do tasks.

context(context)

Old information and rules for work.

Hacker News(Hacker News)

A technology discussion website.

💡 The gist

  • Keep each customer’s rules with the correct project.
  • Save past decisions so helpers do not repeat work.
  • Hacker News attention shows interest, not proof.

A user works with AI agents, computer helpers that can do tasks. The user works across several clients. A client is a customer or organization.

The user says repeating context takes a lot of time. Context means the old information and rules needed for work. This can include chats, emails, tickets, past choices, and customer rules.

The post asks five simple questions. What should enter a new session? How can another agent receive the same history? How can people separate each customer’s information? What happens after an agent forgets a decision? Is there a useful tool to buy or build?

These questions connect to one deeper problem. A new session may not have the old decisions. Then a person must explain the project again. An agent may also miss a plan that people already rejected. If customer information is mixed, the agent may use the wrong rules. The post raises these risks. It does not prove that they happened in this case.

The writer is exploring whether to build a tool. The post does not name a best system. It gives no test results. It gives no measure of saved time.

The thread had 98 points and 58 comments in the supplied snapshot. Points and comments measure attention. They do not prove the solution works. The Hacker News post is the source.

What to watch

The next step would be a shared method or tool. It should show what gets saved. It should show how agents share information. It should also show how each customer stays separate.

For now, the post gives us a useful question. How can AI helpers continue work without forgetting the project?

Source: Hacker News post

💬 How People Keep AI Agents from Losing Context

Most answers keep separate notes for each project or client and let the agent look up what it needs. The numbers and failure stories are self-reports from commenters.

  • Use Markdown or Obsidian notes, design documents, issues, and an index for each project or client. AGENTS.md or CLAUDE.md can point the agent to the right place.
  • At the end of a session, write down what matters in a topic or handoff note. Archive old detail so the active context stays small.
  • Search tools such as QMD can find relevant knowledge. A gatekeeper can require tests, linting, records, and updated documentation before work is marked done.
  • The reported efficiency results differ: LLM Wiki was said to add 23%, graphify 5%, wayfinder to reduce usage by 30%, and another tool to reduce it by 40% in one evaluation. None of these figures was independently verified.
  • Separate accounts, folders, or spaces help keep clients apart. One commenter reported losing a container with sample data after a failure. Others argue that clear Markdown and an index are enough without building a large system.

mature digest at 58 comments (revision 1). We fetched 60 comments and sampled 60 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Can a Computer Helper Remember the Right Job?

📰 Full story: How Can a Computer Helper Remember Each Project?

A person wants help without teaching the same thing again.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
AI agent(AI agent)

A computer helper that can do tasks.

Hacker News(Hacker News)

A website where people talk about technology.

What happened

An AI agent is a computer helper.

It can do jobs for a person.

One person helps several customers.

Each customer has different rules.

The person wants the helper to remember the right rules.

The person also wants old choices to stay with the right job.

A helper might forget a choice.

It might try an old, rejected idea again.

It should not mix one customer’s work with another’s.

The question appeared on Hacker News, a website where people talk about technology.

The thread had 98 points and 58 comments.

Those numbers show attention.

They do not show that the answer is true.

The writer is thinking about making a tool.

The writer has not chosen a tool yet.

We do not know which way works best.

Source: Hacker News post

💬 Give Each AI Job Its Own Notebook

Many people give each job its own notes and folder.

  • Make a separate folder for each project or customer. An index helps the AI find the right page.
  • Write down important things when the work ends. Put old details away. A checker can make sure the tests and notes are ready before saying the work is finished.
  • In one person’s comparison, LLM Wiki used 23% more tokens, graphify used 5% more, and wayfinder used 30% fewer. Another person reported 40% fewer in one test. These were personal reports.
  • Separate customer boxes reduce mixing. Some people still say simple Markdown and an index are all they need.

mature digest at 58 comments (revision 1). We fetched 60 comments and sampled 60 across the thread. These are HN users’ reports, not independently verified facts.

Sources