Adil Karim

ML Engineer & Co-founder @Oiya_ai. Building OpenLore because agent knowledge is infra. Open source.

[Adil Karim @_aakarim]

Your team’s docs shouldn’t live in Git. They should live in a remote filesystem your agents can search.
Launched OpenLore at the Claude Code Copenhagen meetup last week. It serves your docs as a shell over SSH. One Go binary. https://t.co/Rbv3oDHe6o

[Alfie Lambert @AlfieOiya]

We just hit our first 1k stars on GitHub ⭐️

OpenLore started from solving a problem we faced every day (like all good ideas): once multiple AI agents share knowledge, prompts and folders stop being enough.

You need permissions, provenance and rules the agents can’t ignore.

Nice to know we’re not the only ones thinking about this.

If you’re not using OpenLore, there are now 1,000 people whose agents are better set up than yours 👀

https://t.co/bOGblxvL5A

[Alfie Lambert @AlfieOiya]

Want to see which docs are getting bloated, what your agents are actually reading, and where context is wasting your tokens?

OpenLore v0.7.0 shipped last week.

OpenLore can surface the problem, then your agent can fix the structure.

Same knowledge. Less context. Better signal.

Free & open source: https://t.co/bOGblxvL5A

[Adil Karim @_aakarim]

@_chenglou @aidenybai Yikes how did I miss this?

[Adil Karim @_aakarim]

@JoshARosen A lot of paradigms to explore here. We’ve added an automated approval router into OpenLore recently. Seems like it’s just scratching the surface

[Adil Karim @_aakarim]

@airesearch12 Thanks! Could you add GLiNER to this?

[Adil Karim @_aakarim]

@thorstenball Exactly what we’ve put in OpenLore! Even added a separate buffer just so agents can read ahead to suggested files based on the shell history

[Adil Karim @_aakarim]

@rileybrown Try OpenLore, share pretty much anything between your agents. https://t.co/MKhdlUoAe2

[Adil Karim @_aakarim]

@himarkyi Glad to see this is catching on! Exactly the idea behind OpenLore - a knowledge engineering platform. Self hosted and completely customisable.

https://t.co/MKhdlUoAe2

[Adil Karim @_aakarim]

You’re right on it revolutionising context management. We’re using it to tag & search in OpenLore and the performance gain has made our existing evals essentially worthless.

But you would still need some pre-processing on write if you want to throw all the context and do it just in time. We tag documents on the way in per folder and then the search tool gives customised choices for Jev when you search. So it’s folder-orientated constraints.

Jev still charges per token!

[Adil Karim @_aakarim]

Could be both! We use an MCP as a wrapper around a sandbox & VFS. MCP passes plenty of information to handle scoping in the sandbox. All the CLIs in the sandbox are already authed (w/ keycard).

The missing link was having sessions be persistent in the MCP, which was solved in the last major release.

[Adil Karim @_aakarim]

What’s not clear right now is model size, and that meaningfully changes things.

With much faster response times the implication is that either it’s a smaller model or the number of active parameters is smaller. If so, we could see faster, smaller VRAM processing take place in these micro architectures, meaning e.g. retries won’t matter as much since colocation would be feasible.

Apple’s bets on the NPU could have a huge payoff here.

[Adil Karim @_aakarim]

@brandon_galang I wonder how they cap the markdown, must get larger than the context window eventually.

[Adil Karim @_aakarim]

Field guides are a brilliant method for a swarm of agents to document decisions on long-running tasks without conflicts.

All you need is line budgets, and they’re supported natively in OpenLore.

[Kydo @0xkydo]

i didn’t get why cursor made their own git system at first, but after checking out their agent swarm blog, it makes a lot of sense. highly recommend everyone to read it.

they went from doing 1000 commits an hour to 1,000 per second in just a few months. that’s a 60x increase. crazy

https://t.co/RohtiQmI69

[Adil Karim @_aakarim]

@meetgranola Can we set up webhooks programmatically?

[Adil Karim @_aakarim]

@davesnx "Invert, always invert" - Charlie Munger

works for engineering too...

[Adil Karim @_aakarim]

Agents are great at writing… and absolutely terrible at knowing when a document is done.

Engineering plans pulls can pull in marketing context. An architecture doc goes three levels deeper than anyone actually needs (wasting time and tokens along the way!). With multiple agents writing to the same knowledge base, everything bloats.

Every agent that reads those files pays the cost in tokens and quality.

So to fix this, we just shipped context controls for OpenLore.

Now you can set budgets on your files: line limits, token limits, file size caps. You can apply them per directory with glob matching. You can use "sticky" mode to lock a file at whatever size its first write was, or set hard absolute limits.

When an agent writes over budget, the write is blocked. Not truncated or warned, but blocked. The agent has to either rewrite to fit, or create a new file and reference the original. It forces a decision about what actually matters.

This only works because OpenLore controls the write path. With git, you find out after CI. With Notion, you can't find out at all. If you're sharing context across agents and teams, those tools weren't built for this.

OpenLore catches it at write time, because it owns the filesystem.

This release builds on the file history and provenance we shipped in v0.5.0, and is built on our plugin system, so you can define your own write-path rules beyond just size.

OpenLore is a minimal, extensible, agent-native knowledge base that keeps shared context current and inspectable. Open source:

https://t.co/Pe5eRwdP9Q

Try it yourself:
ssh https://t.co/JJKSFEZq3h teach | claude

[Adil Karim @_aakarim]

@dctanner @AlfieOiya @mattpocockuk @mintlify We’ll work on this! When you say internal & external, what external docs would you be showing? API/dev docs or customer support types?

[Adil Karim @_aakarim]

@andrewdsouza 👀

[Adil Karim @_aakarim]

@mikejulian @DuckbillHQ Review with scripts is exactly what we’re doing for OpenLore, but for everything *except* code. It scales better across a team. May be useful for your context poisoning issue.