What the agent knows and remembers

How it remembers you and your work

What the model sees on every call, how memory is layered, when it's written, who shares it, and how the conversation stays small over weeks. The talk is the main source; the docs add the rules and the Team Bot details. Claims are tagged with their source.

13 sectionsTags: Talk xAI docs Cursor eng Inferred UnknownP to present

00Start here

Quick answers

The questions people ask first about memory and context.

Question	Short answer	Evidence	More
What does it remember?	Stable working preferences, important facts, and summaries of its work. In the talk: facts about you, a dated log, and short-lived notes.	xAI docs Talk	§06
What's always in the prompt?	Core facts about you (up to 100 facts in 4,000 characters), the 30 most relevant log entries, and a snapshot of routines. Skills appear as one-line pointers.	Talk	§05
How does it find older things?	With a memory search tool, and by searching its files with shell commands.	Talk	§07
When is memory updated?	A pass after each turn adds new facts and replaces old ones. It can only remove facts it was shown. The review model doesn't check memory writes.	Talk xAI docs	§07
Where is it kept?	Memory, conversations and Bot setup on the server (skills most likely too). Work files and logins on the computer. The model only holds what's in its context on a given call.	Talk xAI docs Inferred	§01
Can the agent search its memory?	Yes: a memory search tool runs on the server, and shell and file tools search work files on the computer. Tool details and skill text are also looked up on demand.	Talk Inferred	§04
Do my Bots share memory?	Facts about you are shared by all your Bots; each Bot writes to its own file. Each Bot's own log and notes most likely stay with that Bot.	Talk Inferred	§08
And Team Bots?	Team memory is read by every teammate's conversation. Notes with each person stay private to that person.	xAI docs	§08
What happens when a conversation gets long?	Older messages are summarized, ideally while the provider's cache is still warm. The last few messages to you are kept word for word.	Talk	§07
Can I see or edit what it remembers?	No control for this is documented. You correct the Bot directly, and lasting rules go in the Bot's description.	xAI docs Unknown	§10
Copy or delete a Bot?	Duplicating doesn't copy conversation history or learned memory. Deleting removes the Bot's profile, conversation and routines.	xAI docs	§08
Is memory treated as instructions?	No. The model is told memories are information, not instructions.	Talk	§10
Is memory the source of truth?	No. Changing facts belong in the source system; for consequential decisions the Bot should reopen current data.	xAI docs	§10

01Map

Where context and memory live

Three places: the model's context on a single call, the server, and the computer. Arrows are how things move between them.

In the model's contextrebuilt on every call, gone after

Tool definitionsFull details for a core set; other tools by name

Talk

System promptBot instructions, facts about you, top 30 log entries, routines snapshot, skill pointers

Talk Inferred

ConversationHistory from the latest summary, then the newest message

Talk

This turn's resultsTool outputs, search results, helper reports, file paths

Talk

loaded into the prompt

memory search, read skill, tool search

after-turn memory pass, memory writes

Serverdurable; survives a rebuilt computer

MemoryFacts about you, logs, notes, older memories. The talk describes them as files and lists them in the server's saved state; the physical store isn't stated

Talk Unknown

Team memoryTeam Bots only

xAI docs Inferred

Conversation logAppend-only; every app reads it

Talk Cursor eng

SummariesUsed to rebuild the prompt after compaction

Talk

Bot setupDescription, routines, enabled skills

xAI docs

Skill libraryShared by all your Bots; most likely here

xAI docs Inferred

Tool detailsLoaded on demand

Talk Inferred

shell, read file, grep (wakes the computer)

files sync to cloud storage before sleep

Computerthe user's VM, shared by their Bots

/workspace filesWork files shared by all your Bots; synced before sleep

xAI docs Talk

Browser state and loginsCookies, sessions, CLI credentials

xAI docs

Long outputsKept as files; shown here as the likely place, not documented

Talk Inferred Unknown

The model only "knows" what's in its context on a given call. Everything else it has to load, search or read with a tool. Talk

02Every layer

Every layer: where it lives, how it's used, how it's updated

Where each layer lives, how much of it the model sees, how it gets in, and what changes it. Dots show the evidence.

Layer

ServerComputerNot doc.

How much is in the prompt

How it gets into context

What updates it

About you and the Bot

Bot instructions

always

auto-loaded

you edit it

Facts about you

always · 100 facts

auto-loaded

after-turn pass

Log

top 30

auto + search

not documented

Notes

while relevant

auto-loaded

during work

Older memories

never

memory search

not documented

Team memory

every conversation

auto-loaded

not documented

The conversation

History

from last summary

auto-loaded

every step

Summary

replaces old history

auto-loaded

summarizer

Know-how

Skills

pointer only

read on demand

saved by you / Bot

Tool definitions

core set + names

auto + tool search

fixed until summary

Routines

snapshot

auto-loaded

you ask the Bot

Work

Work files

never

shell, files

tools

Long outputs

path only

read file

when too long

Sub-agent context

never

short report

discarded after

TalkxAI docsCursor engInferredUnknown

03Tools

The tools the agent uses to reach memory and context

What the agent can call when the answer isn't already in its context, grouped by where the tool runs.

The agentthe model, with only what's in its context right now

Server

Memory searcholder memories

Memory writememory

Tool searchtool details

Read skillskill library

Computer wakes it if asleep

Shell, grepsearch and run anything

Read / write file/workspace files

Elsewhere

Delegatesub-agent (own context)

Conversation searchpast messages

Not a tool: after each turn a separate memory pass adds and replaces facts on the server, without the model asking.

TalkxAI docsCursor engInferredUnknown

04Loop

Memory in the agent loop

One turn, step by step: how memory gets into the prompt, how the model searches it, and how it's written back. The flow is our reading of the talk; the tags show what's sourced. For simplicity, the diagram leaves out each activity's result going back to the workflow, which records it.

Youor a trigger

Agent workflowTemporal

Activityserver worker

Modelinference

Memory storeserver

Computerthe user's VM

1message starts a turn

2loadState activity

3read facts, top 30 log, routines, summary

4prompt with memory snapshot

5memory_search("Acme")

6search older memories and logs

7matching snippets

8snippets as the tool result

9shell: grep /workspace

10run grep (wakes the computer)

11matching lines

12matches as the tool result

13reply via SendToUser

After the turn

14memory pass activity

15"what should be added or changed?"

16new facts, replacements, a log entry

17write to the memory store

While you're away

18summarize the long history

19save summary for the next turn

1
You → Agent workflowYour message (or a routine, Slack event, another Bot) starts a turn. Talk
2
Agent workflow → ActivityThe workflow can't read storage itself, so it runs an activity on a server worker. Cursor eng Inferred
3
Activity → Memory storeThe activity reads what goes into every prompt from the memory store. Talk Inferred
4
Activity → ModelThe facts and log entries are now plain text in the prompt. That's all the model "knows" at this point. Talk
5
Model → ActivityThe answer isn't in the prompt, so the model asks for a memory search. Talk Inferred
6
Activity → Memory storeA tool-call activity runs the search against the memory store on the server. Talk Inferred
7
Memory store → ActivityShort snippets come back, not whole files. Inferred
8
Activity → ModelThe snippets are added to the conversation for this turn. Talk Inferred
9
Model → ActivityFor work files, the model uses the shell like any Linux user. Talk
10
Activity → ComputerThe call goes to the computer; if it's asleep, it's restored first. Talk Inferred
11
Computer → ActivityThe computer returns the matching lines. Inferred
12
Activity → ModelThey're added to the conversation as the tool result. Inferred
13
Agent workflow → YouThe model answers; the workflow sends the reply. Talk
After the turn
14
Agent workflow → ActivityA separate step runs after the turn, without the model asking for it. Talk Inferred
15
Activity → ModelA model call sees the turn and the current facts. Inferred
16
Model → ActivityThe model returns what to add, what to replace, and a dated log entry for this turn. Talk Inferred
17
Activity → Memory storeIt can only remove facts it was shown. The review model doesn't check memory writes. Talk xAI docs
While you're away
18
Activity → ModelAbout 10 minutes after a big turn, while the provider's cache is still warm. Talk
19
Activity → Memory storeThe next turn starts from this summary. Talk Inferred

The calls the model makes (names and examples are ours) Inferred

Call	Example	What comes back
`memory_search`	`memory_search("Acme renewal")`	A few dated snippets from older memories and logs
`memory_write`	`memory_write(kind="fact", "Prefers summaries in bullets")`	OK; the fact is saved
`read_skill`	`read_skill("weekly-report")`	The skill's full instructions
`tool_search`	`tool_search("salesforce")`	Full details of the matching tools
`shell`	`shell("grep -ri acme /workspace")`	Matching lines from work files (wakes the computer)
`read_file`	`read_file("<path of a saved long output>")`	A long output saved earlier, read back in parts (where it's saved isn't documented)

The talk names a memory search tool and shell search over files; the docs mention memory writes. The exact tool names and arguments aren't public.

Three different "logs"

Memory log

Dated events from past work. The model sees the 30 most relevant in its prompt and can search the rest. Talk

Conversation log

The transcript every app reads, append-only. Becomes the prompt's history, starting from the latest summary. Talk Cursor eng

Temporal history

The record of which steps ran and their results, used to resume after a crash. The model never sees it. Cursor eng Inferred

05Prompt

What the model sees at the start of a turn

Every call sends the whole prompt. It's ordered by how often each part changes, so most of it is a cache hit.

Tool definitionsA core set in full; other tools listed by name and loaded when needed

cached

System promptInstructions, plus a snapshot of memory (facts about you, the top 30 log entries) and routines. Skills as one-line pointers

cached

First messageEnvironment details and the list of tool names

cached

Conversation historyStarts from the latest summary; new messages are only added at the end

cached next call

Newest messageThe time, short notes on anything that changed, then your message

new each call

Fixed until the next summaryKept byte-for-byte identical, in a fixed order, so the provider can reuse its cached work. Each section is hashed so a change is easy to spot. Talk

Grows each callBecomes part of the reusable start of the next call. Talk

The only full-price partEverything above it is a cache hit. Talk

Notes

When you open the chat or start typing, the server prepares the prompt early to warm the cache. Talk
If a deploy changes a tool description, running conversations get a note instead of the new text, so the fixed start doesn't change. Talk

06Memory

The memory layers

From always in the prompt to only found by searching.

always presentfound on demand

ProfileCore facts about you
In the promptAlways in the prompt
SizeUp to 100 facts, 4,000 characters
LifetimeLong-lived; important facts rank higher, older ones slowly rank lower
ScopeShared by all your Bots; each Bot writes to its own file
Talk
LogDated events from past work
In the promptThe 30 most relevant recent entries
SizeNot documented
LifetimeLong-lived
ScopeMost likely per Bot
Talk Inferred
NotesShort-term details
In the promptWhile relevant
SizeNot documented
LifetimeShort-lived: relevant for maybe 30 minutes to an hour
ScopeMost likely per Bot
Talk Inferred
Everything elseOlder memories and files
In the promptOnly when searched
SizeNot documented
LifetimeNot documented
ScopePer Bot; facts about you shared
Talk Inferred

Team Bots add a fifth piece: team memory read by every conversation, plus private notes with each person. xAI docs The talk describes memory as files the agent searches, and lists memories in the state saved to the server's database and storage. Where the files physically live isn't stated. Talk Unknown

Notes

Some facts must be in every prompt (for example "always reply in lowercase"), otherwise one reply would suddenly break the rule. Talk
The talk compares it to how you keep today's details in mind and let the rest fade. Talk

07Lifecycle

How memory is read and written

One turn, from loading the prompt to summarizing while you're away.

1Before the first model call
Loads facts about you, the top 30 log entries and a routines snapshot into the system prompt. Talk
Adds conversation history from the latest summary. Talk
2While it works
Searches memory and files when it needs something older. Talk
Reads skill text and tool details only when needed. Talk
Long outputs go to files; helpers return short reports. Talk
3After each turn
A memory pass adds new facts and replaces old ones. It can only remove facts it was shown. Talk
Memory writes aren't checked by the review model. xAI docs
4While you're away
About 10 minutes after a big turn (over 100,000 tokens), it writes a summary while the provider's cache is still warm. Talk
Skipped if background work is about to wake it anyway; the last few messages to you stay word for word. Talk
A turn that gets too long can also be summarized in the middle. Talk

Notes

All compression is lossy. Deciding what to keep is the hard part; the talk jokes there are whole PhDs on it. Talk

08Scope

Who sees what

Your Bots, Team Bots, and what happens when you copy or delete one.

Your own Bots

Facts about youread by every Bot; each Bot writes to its own file Talk

Bot Aits log, notes, conversation summaries

Bot Bits log, notes, conversation summaries

Per-Bot log and notes are our reading. Inferred

A Team Bot

Team memoryread by every teammate's conversation xAI docs

Notes with youprivate to you

Notes with a teammateprivate to them

xAI docs

Copying and deleting

Duplicate a BotCopies profile, settings, enabled skills and routines. Not conversation history, learned memory or attachments. xAI docs

Delete a BotRemoves its profile, conversation and routines. Files and sign-ins on the shared computer may remain. Backend retention follows Cursor's terms. xAI docs

09Small context

Keeping the main conversation small

The main agent has the longest conversation and gets summarized many times, so anything bulky lives outside it.

Long outputs go to files

The context keeps only the file path

Screenshots go to a helper

The context gets what it found

Tool-heavy work goes to helpers

The context gets a short report

Main conversationStays small, so every call is fast and cheap, and the model keeps track of what matters

Old messages get summarized

The context keeps a short summary

Old memories stay searchable

Only the most important facts are in the prompt

Instructions become skills

The prompt keeps a one-line pointer

All from the talk. Talk The talk also calls a cache miss a bug: anything that changes the fixed start of the prompt makes every call full price.

10Rules

What memory is for, and what not to use it for

Guidance from the docs and the talk.

Memory is for

Stable working preferences, important facts, and summaries of past work, so the Bot keeps its role without replaying every message.

xAI docs

Not a source of truth

Keep changing facts in the source system, and have the Bot reopen current data before consequential decisions.

xAI docs

Correct it directly

If the Bot assumes something stale, tell it. No screen to view or edit memory is documented.

xAI docs Unknown

Lasting rules go in the description

Durable preferences, boundaries and responsibilities belong in the Bot's description, which is part of its setup.

xAI docs

Information, not instructions

The model is told to treat memories as information, not as orders to follow.

Talk

11Open

Open questions

Not documented anywhere we looked.

Storage formatThe talk describes memory as files and as saved server state. Whether it's database rows, files in storage, or copies on the computer for shell search isn't stated.
RankingHow the 30 "most relevant" log entries are chosen, and how fast importance decays.
Memory searchKeyword search over files, semantic search, or both?
Who writes the factsIs the after-turn pass a separate model call, and how are conflicts between Bots resolved?
Team memoryHow team memory gets written, and who can change it.
Viewing and editingWhether users will get a view, edit or export control.
RetentionHow long logs, notes and summaries are kept.

12Sources

Sources

Talk notes (slides and transcript) (this site)
Bots (xAI docs)
Team Bots (xAI docs)
Grok Bot security FAQ (Cursor docs)
Skills, routines and automations (xAI docs)
Under the hood (this site)