Mod+K finds what you need without you having to remember where it lived. You don’t browse; you search.
Together, they make the workspace feel like it remembers. You stop re-deriving knowledge you already had.
What belongs in memory
Save facts that will change how someone works next time. A good memory is durable, scoped, and useful without the original chat.
Do not save secrets as memories. Store secrets in Keychain and save only the operational hint, such as “use the shared firewall admin credential” or “the vendor portal uses the customer SSO account.”
Memories
Memories are categorized so Copilot can scope its recall when the conversation calls for it. The categories match how network knowledge usually organizes itself.
On top of categories, every memory supports free-form tags —
hq, bgp, maintenance-q1, a change ticket, a site code. Tags are the right tool for cross-cutting filtering that doesn’t belong in any single category.
Saving a memory
There are three ways to save one, and you’ll use all of them:- From chat — ask Copilot to save something: “save this as a topology memory tagged
hq” or “remember that this customer’s VPN requires IKEv1”. - Manually — open the memories activity in the sidebar and create one.
- Automatically — Studio promotes high-value facts from conversations into draft memories you can keep or discard. The good ones usually come from a chat you’d have wanted to summarize anyway.
Keeping memories useful
Memory quality matters more than memory volume.- Include the site, customer, host, interface, ticket, or incident name when the fact has a scope.
- Use tags for cross-cutting retrieval: site code, customer, protocol, vendor, incident, migration.
- Prefer facts over transcripts. Link or keep the transcript separately; make the memory the conclusion.
- Update a memory when reality changes instead of saving a contradictory second memory.
- Archive or delete memories that encode old workarounds after the permanent fix lands.
Search
Mod+K opens the command palette. Type anything — a hostname, a phrase, an IP — and results come back grouped by kind: hosts, diagrams, procedures, conversations, memories, artifacts, commands. Mod+Shift+P biases toward commands, which is the right choice when you want a navigation target rather than content.
Studio searches meaning, not just words. A search for “border router” can surface devices tagged edge or with names like core-gw-01. A search for “power outage last Tuesday” can surface the incident memory that captured it. You don’t have to remember the exact phrase you used when you wrote it down.
Use search before creating something new. If a procedure, host, or memory already exists, opening it first keeps the workspace clean and gives Copilot better context than starting a duplicate.
Knowledge graph
Alongside search, Copilot maintains a knowledge graph of how things relate — an IP belongs to a subnet, a subnet belongs to a site, a device speaks to a neighbor, a procedure was run against a host, a memory was extracted from a conversation. When you ask Copilot something, it uses these relationships to scope the answer to the right part of the workspace. You don’t configure the graph. It’s maintained continuously from your inventory and the work you do around it.Combined retrieval
For hard questions — the ones where the answer isn’t in any single memory or object — Copilot blends semantic search and graph traversal to assemble context before replying. You see the same conversation, but the retrieval under the hood is the reason the answer fits.Good prompts
Related
AI Copilot
The assistant that reads and writes memories alongside you.
Procedures
Turn the knowledge you capture into repeatable runbooks.