[ shared wiki ]

Your Aeris agents share a brain.

Andrej Karpathy described an "LLM Wiki": instead of re-reading everything from scratch every chat, the system compiles your raw sources into a real, interlinked wiki of markdown files that grows over time. Aeris does exactly that for your agents.

Every agent writes to the same fabric. Aeris reads each write, extracts entities and topics, and updates a persistent wiki of plain markdown pages. The wiki lives as a folder of .md files — open it in Obsidian, commit it to git, grep it from the terminal.

Watch the wiki build itself

Write as any agent. Entities and topics appear in the sidebar. Other agents can recall them through wiki_ask.

fabric_write

writes →

wiki_ask

ask anything any agent has written.

entities/

topics/

~/fabric/wiki/entities/Northwind.md

Northwind

aeris-01 · #1

Vendor Northwind requires PO before invoice; net-30 terms.

cross-references

Entity · VendorEntity · POEntity · net-30Topic · requiresTopic · invoiceTopic · net-30

The explorable graph

Every write becomes a node. Entities and topics are the cross-references. Hover a node to highlight its edges.

wiki graph · 25 nodes · 23 edgeshover a node

entitytopicwrite

shared fabric status

all agents connected

writes

entities

topics

agents

who's writing

aeris-011 writes shared
aeris-021 writes shared
aeris-031 writes shared

last cross-agent recall

aeris-03 wrote: "Acme contact is Priya at priya@acme.co; she signs purchase orders."

other agents can now use wiki_ask to retrieve this.

Command console

Run wiki_llm_status, fabric_write, and wiki_ask as any persona. Output goes straight into the same fabric as the panels above.

aeris/console

personaagent

aeris console · persona=Mira Ahn agent=aeris-01

type "help" for commands

Training data export

Best writes get exported as JSONL fine-tuning pairs.

training data export

min quality score0.40

{
  "messages": [
    {
      "role": "system",
      "content": "You are Mira Ahn, an Aeris agent. Answer from shared fabric memory when relevant."
    },
    {
      "role": "user",
      "content": "What do we know about Vendor?"
    },
    {
      "role": "assistant",
      "content": "Vendor Northwind requires PO before invoice; net-30 terms."
    }
  ],
  "meta": {
    "source": "aeris-01",
    "id": 1,
    "score": 0.68
  }
}

{
  "messages": [
    {
      "role": "system",
      "content": "You are Tev Okoro, an Aeris agent. Answer from shared fabric memory when relevant."
    },
    {
      "role": "user",
      "content": "What do we know about Customer?"
    },
    {
      "role": "assistant",
      "content": "Customer Acme prefers webhook delivery over polling for order events."
    }
  ],
  "meta": {
    "source": "aeris-02",
    "id": 2,
    "score": 0.7
  }
}

Each row is a {messages, meta} sample — drop straight into your fine-tuning pipeline.

Shared memory across all your agents

One agent learns it, every agent can recall it.

A wiki that builds itself

Every write gets extracted into entity pages, topic pages, and cross-references.

Memory maintenance

Quality scoring, duplicate detection, auto-archival of dead weight. Your brain stays clean as it grows.

Training data export

Your best agent work becomes fine-tuning pairs for replacement models.

get the update (one command)

curl -fsSL https://aeris.dev/install.sh | sh

try it in your Aeris chat

use wiki_llm_status — confirm the model is wired up
use fabric_write to record any decision, fact, or fix
watch ~/fabric/wiki/entities/ and ~/fabric/wiki/topics/ fill in
use wiki_ask "what do we know about X" — synthesized answers with citations

Every future session starts with that knowledge already loaded. That's the whole pitch.

Inspired by the Icarus wiki plugin and the Nous Research Hermes agent ecosystem.

Try the decade simulation →Read the manifesto