Perplexity built a personal wiki for its AI agents

If you have been working with an AI agent for months across hundreds of sessions, having to remind it that you prefer PDF exports over PNG gets old very quickly.

Most agents today either forget everything between sessions or try to cram their entire history into the context window until the model falls apart. On August 19, Perplexity unveiled Brain, a structured memory architecture designed for its Computer AI agent that organizes long-term context like an internal knowledge wiki.

A three-layer Markdown filesystem

Instead of building exotic memory databases, Perplexity decided to treat memory as plain files inside the agent's sandbox.

Brain organizes persistent context into three distinct filesystem tiers. The sessions directory preserves raw transcripts and summaries. Above that, the notes directory stores distilled snippets organized by topic.

At the top sits the knowledge directory, which functions as the actual Brain. It synthesizes insights across sessions into interconnected Markdown pages.

These pages rely on two specific types of connections. Wikilinks connect concepts and projects horizontally, answering what else the agent needs to know. Citation tags link claims downward to raw transcripts, allowing the agent to verify whether a fact is actually true.

The entire wiki is backed by Git, preserving diffs and change logs so multiple subagents can read and update memory at the same time without overwriting each other.

Skipping the 500x latency trap

Navigating persistent memory is notoriously prone to latency bottlenecks.

Perplexity initially considered mounting the entire history as a remote filesystem. However, agentic workflows often run thousands of filesystem commands in a single sweep.

Internal tests revealed that running a simple grep across a remote FUSE-backed filesystem was 400 to 500 times slower than doing it locally, grinding agent exploration to a halt.

To fix this, Brain relies on a hybrid materialization strategy. When an agent boots, it loads a compact index and an initial working set of recent files.

If the agent needs deeper context, it spins up a specialized Memory Agent to run semantic search. The relevant records are then materialized as local files on demand, keeping subsequent grep and read operations lightning fast.

Why it matters

Persistent memory has been one of the messiest engineering problems in the agent space. Cramming static logs into prompt windows burns money, while vector databases frequently hand the model isolated, decontextualized fragments.

We think Perplexity made a brilliant engineering call by doubling down on plain text files, Git versioning, and standard shell utilities. It turns the agent's memory into something human-readable, auditable, and easily traversed with basic CLI commands.

For power users tackling long-running research or complex multi-week coding tasks, this could finally make an AI assistant feel like a colleague that actually remembers past decisions without needing constant hand-holding.

The potential risk, in our view, lies in the synthesis layer. If the background agents hallucinate or misinterpret a past session during the distillation process, that error gets baked into the wiki and cited in future sessions.

Still, replacing opaque retrieval black boxes with an inspectable Markdown wiki is one of the most practical approaches to agent memory we have seen so far.