This is actually useful. LangChain just released OpenWiki. It's an open-source agent that creates a wiki for your codebase, connects it to your coding agent, and keeps it updated as your repo changes. Your AI coding agent gets long-term repo context without stuffing everything into CLAUDE.md. Here's how to... setup. Save this.show more

GitHub repo:

Wait isn’t there gitingest already?

Surprised it’s not called LangWiki. Feels off brand for them, not sure why…

Cool stuff. This is a limitation in today’s agents. I would love something similar but for learnings, workflows, user common preferences, etc … say my agent once spent 5 cycles attempting something until it “learned” what was blocking it and succeeded at the task, then the wiki grabs it as a learning so next time the cycle is faster. I know this can be done with Karpathy’s wiki approach (I’m already doing it). However just kinda weird why coding agents like @cursor_ai @claudeai and codex dont do it out of the box.

Been solving this manually for close to two years now, back when context windows were tiny and models were nowhere near this capable. My approach: Before any code gets written, I build a foundational documentation set. Architecture file, project blueprint, UI/UX scenarios, features definition. All markdown, all written before an agent touches anything. During the actual build, I feed the relevant files into context as needed instead of dumping everything at once. I maintain running memory through AGENTSmd and CLAUDEmd, plus a separate memorymd file. And once I hit around 80% of the context window, I write a handoff file, basically a compressed briefing that lets a fresh agent session pick up exactly where the last one left off without burning tokens re-deriving everything. It works, but it's genuinely a lot of manual upkeep. That's exactly why I'm glad something like OpenWiki exists now, it automates a big chunk of what I've been doing by hand for codebases specifically. One more piece I've added recently: I run this whole memory layer through The Curator, an app with an MCP server built in, so agents can read and write to a persistent knowledge graph instead of me managing every file manually. Feels like one more piece of the same mosaic everyone's converging on. Agents need an actual memory layer, not just a bigger prompt.

been looking for something exactly like this the CLAUDE.md approach was getting messy fast

Nice

Is this line a complement to, or substitute of, graphify? Anyone tried using both?

Is it possible to let the coding agent update openWiki? So it can always stay up to date, no need for another model api, and coding agent might have clearer context on what is the change and what worth remember.

the real bottleneck was never context, it was nobody wanting to babysit a wiki. this just automates the babysitting

the auto-updating part is the whole point. a hand-written codebase wiki rots the second someone refactors, and a stale wiki is worse than none - the agent trusts it and confidently builds on context that's no longer true. keeping it in sync is the hard problem, not generating it.

This is useful 👍🔥 Thanks for sharing

Careful... Medusa flagging 13x Critical concerns incl Security issue detected: agentic-supply-chain-tool-compromise

I love to see this as a standard. Ideally though, this would be baked into harnesses not requiring a 3rd party CLI / subscription.

¿Y el contexto operacional? Permisos, decisiones históricas, por qué un componente está aislado. Eso no emerge del código.

LangChain's OpenWiki is a smart move for agent memory. We've seen similar context wins in Supabase + vector stores for our couple app — the agent can now reference past decisions without re-summarizing the whole repo every run. Curious how it handles merge conflicts in the wiki.

the keeping-it-updated part is where this lives or dies. if the wiki drifts from the actual repo state, it's worse than no context because the agent trusts stale info. wonder how it handles merge conflicts and rebases

so now we just get a wiki instead of the usual waitlist spam? guess ai is finally reading documentation too

I solved it using a default CONTEXT.MD file for each folder like a init file and rules file or folder at root level that takes care inter-relationships. I ask agent to update relevant context files after every valid change.

Wait, so no more bloated CLAUDE.md? Finally. This actually solves a real problem. Saving this one.

LangChain Really wanted to Build the Entire Agentic Ecosystem, with their Products catalogue its quite clear.😅

Oh, so I don't have to stuff my repo into CLAUDE.md anymore? About time documentation evolved with the agents writing it.

How’s this compare to GitNexus?

repo-wide context without the clutter is a game changer

been looking for something exactly like this

Massive workflow upgrade.

tried the 'dump everything in context' approach → broke at ~200k tokens. selective file injection works better for me. how does this handle cross-package references in monorepos?

I need a quick reference for all of yalls acronyms lol

That's a solid find .... Gonna try today 😸

Same staleness problem, code wiki or notes. Answer to Shrey's either/or: neither. Trigger the update off the agent's own read/edit during a session, not a commit hook or a diff heuristic. Wrote about the personal-system version:

The repo wiki is useful only if it stays close to the diff. A stale explanation is worse than no explanation because the agent trusts it with confidence.

How is this different than knowledge graph?

it only lets you use an API key to generate this?

Stale docs are the silent killer here. A wiki that updates as the repo changes fixes the trust problem, most internal docs die because nobody believes them after month two. Curious how it handles conflicting context between branches.

Cool! Super helpful.