Memory for agents is still early, with little to no standards. “Memory” means something different to everyone. But one common pattern is emerging: wiki memory.
The idea is simple: use an agent to turn raw source data into a compact, persistent, agent-readable knowledge layer.
Why wikis?
Raw data contains a lot of knowledge, but it is often inefficient to expose directly to an agent. Logs, notes, code, docs, experiments, Slack threads, and transcripts are too noisy and too large. So instead, we run a process over that data and transform it into a denser representation.
This is different from basic RAG. RAG usually retrieves raw chunks at query time. A wiki precomputes and maintains a higher-level synthesis, so the agent does not have to rediscover the structure every time.
This desire exists almost everywhere. When speaking to a friend at a research company, he talked about all the knowledge in their researchers' heads. He wanted to “clone their brain” so that if they left, the knowledge still remained with the company. His hope was that by looking at the experiments they ran, the notes they wrote, and the actions they took, they could approximate this “brain clone.”
A wiki is one practical way to do that: not by storing everything, but by compressing what matters into a reusable knowledge base.
What is a “wiki”?
A wiki is an agent-maintained data structure that represents source knowledge in an agent-friendly way.
In practice, this often means running an agent over some source material and asking it to create a set of files that future agents can use to understand the domain faster.
The important bit is not that it literally looks like Wikipedia. The important bit is that it is persistent, structured, inspectable, and updated over time.
Examples of wikis
DeepWiki by Cognition is probably the first example of this I remember seeing. DeepWiki creates AI-generated documentation for GitHub repositories. It is intended to give humans and coding agents a higher-level mental map of a codebase, making it easier to understand and navigate.
Karpathy recently wrote about what he called an “LLM Wiki” or “LLM knowledge base.” This is a more general form of the same pattern: instead of only working over code, it can work over arbitrary source files. His framing is that the LLM incrementally builds and maintains a persistent markdown wiki that sits between the user and the raw sources.