LangChain CEO Spotlights Wiki Memory Pattern in AI Systems

@hwchase17 @cognition @FactoryAI @karpathy wiki memory is great right up until the wiki is confidently wrong and now every agent inherits the same bad assumption. storing it was never the hard part, trusting it is.

this may be ai generated, but is true - the hardest part of wikis (and memory in general) is the process that condenses raw data into learnings/memory

Memory has somehow consistently been the most exciting area of agent development over the last 3 years (imo), and it's still a largely unsolved problem!!

Wiki's are the biggest advancement I've seen in a long time, and I'm super excited to push more on research here. Great article on them from Harrison, and stay tuned for more memory + wiki OSS from LangChain in the coming days!

human in the loop is very nice for these wiki memory systems

good description of how someone has done wiki memory in the wild

Wiki wiki wiki wiki

@hwchase17 I want a metaphor similar to how a grain of sand forms a pearl, but with a negative outcome…

@hwchase17 @cognition @FactoryAI @karpathy Files are fine for local but they break down when you want to go Enterprise level with security, permissions, concurrent writes by multiple agents etc. Agree or disagree and if you disagree what is your opinion on the way files work in enterprise?

@hwchase17 @cognition @FactoryAI @karpathy Wiki memory feels useful because it turns memory from “more context” into an operational artifact.

For production agents each entry needs provenance, freshness/expiry, write permissions, and rollback path.

Otherwise memory becomes another silent prompt injection surface.

@hwchase17 I’d keep memory proposals in review until they show up in real work twice. One bad accepted fact can make every later run look confident and wrong.

@hwchase17 @cognition @FactoryAI @karpathy this setup for persistent memory actually makes so much sense

I hit this with my Claude Code wiki experiment. file_write hooks captured the diff without the reasoning, and the session-end compiler kept recycling the same open questions. The fix is probably keeping the raw trace plus a short memory card for status and evidence, with the wiki page reduced to an index.

@hwchase17 I was gonna say, I doubt it's AI, but that last sentence was sus. 😭

Nevertheless, agreed.

@hwchase17 Hosting files as doc db objects with an API and CLI for agents is clean - it externalizes state without bloating the agent's context. Been hitting these same design questions building multi-agent systems in LangGraph. Retrieval process matters more than which db you pick.

@hwchase17 totally, it makes the whole experience feel so much more grounded

@hwchase17 that approach to persistent memory is actually genius

@hwchase17 Agreed — the practical bottleneck is how often humans actually review vs. rubber-stamp the suggested recall.

@hwchase17 Memory quality lives in the distillation pass.