It remembers

Tell it something once, in an ordinary chat:

Remember that we invoice on the 20th, and we never chase before the 25th.

Weeks later, in a brand new chat, ask what to do about an unpaid invoice. You get the 25th back without repeating yourself. The note sits on the Memories page (Manage > Memories, /manage/memories), where you can read it, edit it or throw it away.

That is the difference from a chat window that starts blank every morning.

Agents look things up before they start

Every agent runs the same short loop whenever work begins or changes direction: search what is already known, do the work, save anything worth keeping. That search covers your memories, past chats, jobs and tasks, files, and your agents' skills.

So when you hand a new agent a quote to draft, it finds the pricing decision you made in March before it writes a word. You never trigger it yourself.

What you can ask it to remember

• A fact or a preference. "Remember that Hanley pay net 30." It lands in a memory and appears on the Memories page straight away.
• A whole body of knowledge. "Make a place to remember our refund policies, and our support bot should be able to read it." Player One asks what it holds and who may read it, then creates it. No folders, no categories, no form.
• A pile of documents. Hand it your service agreement, price list and onboarding pack. Source files are kept whole rather than summarised, and become searchable by meaning.
• A way of working. "Always write to me in British English and keep replies short." Standing instructions like that attach to the agent and are read on every turn, so keep them few and short.

To fill a new knowledge base, say so and Player One will sit with you: one question at a time, you answer or paste, it files each piece as you go. Build a knowledge base has the steps.

Where it keeps things, and who can see them

What Player One remembers is grouped into Memory Spaces. Each agent has one of its own. You create shared ones for anything the whole team should know: a support knowledge base, a client dossier, your internal wiki, an archive of documents.

Opening a space shows what it is for, what belongs in it, the files themselves and its search index health. Two settings matter to you. Audience is either internal, meaning you and your agents only, or public, meaning a customer-facing bot may answer from it. Player One warns you before creating a public one, because a public space can be read outside your installation. Sharing decides which projects a space is connected to, and whether agents there may only read it or also add to it.

A customer-facing bot only ever reads. It cannot write to your memory, so nothing a member of the public types can be planted there; anything it notices comes back to you as a suggestion.

Teaching it, so the lesson sticks

Correct an agent once and the correction is recorded against the skill that governs that kind of work. The next time any agent does that work, the correction arrives with it. You do not correct the same thing twice.

When a whole way of doing something is worth keeping, say "turn that into a skill". Player One works through it with you: what the skill should produce, when it should fire and when it should not, then trial phrases to check it fires at the right moment and stays quiet otherwise. You choose whether it belongs to one agent, one project, or everyone. Skills are listed under Admin > Skills (/admin/skills).

More in Teaching it and Skills.

Finding anything you have ever done

The search box at the top of every page searches your chats, agents, jobs and files together, opening results at Search (/search) where you can narrow to one kind. A chat hidden from the sidebar is still findable there.

Or just ask an agent: it searches the same material and answers from it rather than handing you a list.

What "gets better the longer you use it" actually means

Three things work away underneath, and one page explains them properly: Memory, explained.

Fold is the index that lets an agent search by meaning, so "what did we agree about late payers" finds your note about chasing invoices. Engram is the curator that tidies your Memory Spaces, filing what has been dropped in and merging it with what was already there. Sleep is when that tidying happens: a background pass on a schedule, not something you wait for.

A new note is saved and visible at once, searchable as soon as the index catches up, and properly filed after the next sleep pass. If you need it usable immediately, ask for a sleep pass and one is queued.

The compounding is that plain: every correction, every document and every decision is one more thing the next agent finds before it starts.

Honest limits

• A memory only recalls well once its search index is attached. New spaces get one automatically, verified before Player One tells you the space is ready.
• Sleep runs to a schedule rather than after every save: by default, fifteen pending items or two days, whichever comes first.
• Rebuilding a space's index from scratch costs real money in tokens, so Player One shows the size and asks you to confirm. Indexing only what changed does not.
• Pausing indexing stops the background work, but searching still costs a little, so it does not stop all spend.
• Spaces made before the current layout look different on the page. They work exactly as well; converting one is tidying, not repair.

Where to go next

• Memory is the short description of the part itself.
• Memory, explained covers spaces, Fold, Engram and sleep in full.
• Teaching it takes a conversation through to a finished skill.
• Create a support agent puts a knowledge base behind a bot.