introducing XERJ Corpus Hub, reference memory for AI agents. give your agent a corpus for your task. watch it become superintelligent. > finds the answers in indexed expert knowledge > applies them to your code, research, and projects > works with any model. free and OSS forever/Apache 2 install... in one prompt: >Install XERJ (docs: index this project’s sources, docs, and logs, and connect as reference memory hubshow more

we believe the future belongs to specialized models. XERJ now allows any LLM to be specialized with public reference memory Corpus Hub, without retraining. Now even small models can compete with frontiers having cloned repos and external memory as a cheat-sheets.

your agent shouldn’t have to rediscover the same knowledge on every task. you can see how it works in the video. today, when an agent needs information, it uses legacy search based on grep, released back in 1974. it reads through files, spending time and tokens looking for answers. along the way, it may also encounter confidential information in files it can access. Xerj is an open-source search engine built for agents. it indexes the files in advance and returns the relevant passages when your agent needs them. the same model becomes: • 10× faster • 20× cheaper • more accurate

agentic search is everywhere. coding: find the exact definition in a codebase the model has never seen. law: find the right clause across 418 UK acts. the challenge: navigating thousands of files and 21,000 sections. this demo shows how indexed knowledge gets your agent to the exact answer while it would otherwise still be searching.

we love how this demo shows what the right knowledge reference adds to product development. 2× faster. 3× cheaper for the same or even better quality, especially for non tier-1 models

give your agent the knowledge it needs. pick a ready-made corpus or build one from your own projects. free. open source. works with any model. get started →

Does this work with anything other than codex and Claude? Can I use it with any agent?

yes, it’s model-agnostic. any agent that can call Xerj’s search tools can use the same corpora, not just Codex or Claude. your data stays with you, and you can keep using the same corpora when you switch models or agents.

Context engineering is mostly a retrieval problem now. Tools that help agents quickly find the right knowledge will matter a lot.

agreed. getting the right passage into context matters more than dumping another thousand pages into it

Honestly, better retrieval can solve more problems than people think.

I can NOT agree MORE! And this is not only about retrieval itself, and how much tokens you burn for that. This is also about raw data->AI data transformation where complex structures like code/AST or relations knowledge like legal case->law became structural for AI, not plain text, not vectors.

This looks like a tool worth exploring for my own projects. Excited to see what people build with it, and congratulations on the launch!

thank you! curious what you'd try it on first

The interesting part is giving agents access to the right knowledge, not just a more capable model. Better context means less time searching through sources and more time actually solving the problem.

you nailed it. a lot of “thinking” is really just figuring out where to look.

Giving AI agents access to indexed expert knowledge could make them far more capable and useful.

working examples make a huge difference. give it something good to learn from

Nice launch, reference memory for agents is a powerful idea

making expert knowledge accessible to any model is a big plus

and you can keep the knowledge when you switch models

how often does the index update when the codebase changes?

incrementally, check this --watch flag we did it right

Memory for AI agents? I’m in man

that what we do

half of this is just pointing your agent at docs

yes, the questions is how much tokens you will waste for 1000 docs :) and how much time

nah that part is actually your problem to solve

not only, imagine you have structured data like legal docs with relations to laws. Or code with cross-references where you need to parse it to AST. XERJ autoindex it and agents knows HOW it's possible to search for such things. All with no tokens at all

no no i'll give you that one tokens are the only part of the brain anyone uses rn

You’ve got a solid idea here, especially for teams building agents around their own data.

which is every team in fact

Open source makes this even more interesting.

yes

beats my rag setups, connecting any model is pure win

curious what you tested it on — what was your previous setup?

small model + good reference memory is how I'd want to run cheap agents. one question before I index my logs: does anything from my local corpus go up to or does the hub only send knowledge down?

no Indexing and search run locally - nothing from it is uploaded to Corpus Hub.

Specialized knowledge without retraining. XERJ is onto something!

exactly. the knowledge is already out there - your agent just needs access to it

My team has been using xerj for two weeks now. It really changes everything.

your corpus is really interesting, thanks for that

You’ve got a solid idea here, especially for teams building agents around their own data.

exactly. your own code, docs, and past decisions are often the most useful context you can give it.

Thanks! Will test

share your feedback pls

This could save developers a lot of time searching through documentation.

@techxutkarsh and save the agent from reading the same docs over and over 😅

Giving AI agents access to structured, expert knowledge can make them far more useful.

especially knowledge specific to the task. “everything about coding” is a lot less useful than a good reference for what you're building

this actually makes a lot of sense for agents

yeah, having them rediscover the docs on every task never made much sense to me

External memory could make smaller models much more capable.

rfight indexed knowledge is a next step to agi

this makes AI agents a lot more useful.

and cheap

curious how much this cuts repeated token usage on bigger repos.

up to 70% of output tokens

This is brilliant giving any agent superintelligent reference memory with indexed expert knowledge in one prompt is a game changer!

would love to see what you try it on!

I’m looking forward to testing this out in my own projects. Congrats on the launch excited to see where this goes!

thx

Reference memory is the missing layer — most agents recompute context they already had. How do you handle invalidation when the underlying corpus changes?

this is the kind of tool i want to experiment with on my own projects. huge congrats on the launch!

thanks, not the last launch at all, again, it is all OSS Apache 2, free forever for everyone

Keeping project notes and docs in one place sounds useful.

you also need an index search for it