Integrations

This page covers two things. How to bring your team's conversations and documents into a memory, and how to give AI tools and agents controlled, audited access to read it back out.

How information gets in#

A memory fills itself from the places your team already works. Everything arrives through the same pipeline. Chatter is dropped, facts are extracted with their sources, and trust rules are applied on the way in.

• Add knowledge in the app. Drop documents onto the Knowledge page, several at a time. Word, Excel, PowerPoint, PDF, OpenDocument, and text files all work. The memory names each one and shows you what it picked up.
• Outlook. Your inbox and your own sent mail are read, so what your mail service files as junk never reaches the memory, while the commitments you make in your own replies are remembered. Quoted reply chains, signatures, and app footers are stripped first, and machine mail is ignored entirely, from notifications and meeting responses to newsletters.
• Outlook Calendar. One more approval on the same Microsoft sign-in and your meetings are remembered too. Who met, when, and what it was about, filed into the rooms they fit.
• Teams. Choose the channels and chats to read. Messages are read as conversations, so a decision that forms across five replies is remembered as one thing. When transcription runs in a meeting you organize, the transcript is read with each line attributed to its speaker.
• Google Drive. Pick folders and their documents sync in. Editing a document replaces the facts its old version created.
• OneDrive and SharePoint. The same Microsoft sign-in covers your files. Pick the folders to read, including ones shared with you, with the same freshness rule as Drive.
• GitHub. Approve the Spark GitHub App and pick repositories. Readmes, docs, and the files you choose are read as documents and kept fresh as they change.
• Discord. Add the Spark bot to a server you manage and pick the channels to read, or read them all.
• Webhooks and the API. Anything that can POST can be a feed, from CI systems and ticketing to telemetry and meeting transcripts. Each delivery is filed into the room it fits, or into the one room you pinned the feed to.

Every source is read only. Feeds are managed on the Sources page, where each one gets its own key, its own trust level, and the rooms it may write to. A feed that turns bad can be paused, and everything it wrote can be removed by name without touching anything else.

Connections belong to the person#

A memory belongs to one person, so its connections are that person's connections. Your memory reads your mailbox and your folders, and a colleague's memory reads theirs. Nothing is shared between memories except what the owners choose to share, and removing one person's connections touches only their memory.

Microsoft 365#

One Microsoft sign-in covers Outlook, Teams, Calendar, transcripts, OneDrive, and SharePoint.

1. On your memory's Sources page, open Outlook or Teams and choose Connect.
2. Sign in with your Microsoft account on Microsoft's own screen and approve read access. Each surface asks only for what it reads.
3. Pick the inboxes, teams, chats, calendars, or transcripts to remember, by name, and the room each one files into.

The first read brings in recent history, so answers work from day one. Reading Teams channels needs a one time consent from your Microsoft administrator, which Microsoft asks for on the same screen.

Google Drive#

1. On your memory's Sources page, open Google Drive and choose Connect.
2. Sign in with your Google account on Google's own screen and approve read only access.
3. Choose the folders to sync and the room each one files into.

GitHub#

1. On your memory's Sources page, open GitHub and choose Connect GitHub.
2. Approve the Spark GitHub App on GitHub's screen and pick the repositories it may read.
3. Back in your memory, choose which approved repositories to remember.

You can change the repository selection at any time from GitHub settings or the Sources page.

Discord#

1. On your memory's Sources page, open Discord and choose Add to your Discord server.
2. Approve the Spark bot on Discord's own screen, on a server you manage.
3. Pick the channels to read by name, or turn on reading every channel so new ones join automatically.

The bot sees only the channels its role can view, so anything you hide from it in Discord stays hidden here too. Removing it from the server ends the connection from Discord's side at any time.

Power Automate and similar tools#

Import the connector file from the Sources page. The connection key is the workspace id, then a dot, then the feed key. Every flow gains a Remember this action that files what it sends into the room it fits.

Trust is enforced on the way in#

Every feed has a trust ceiling, the best source rank its records can ever claim. Chat and email stay at the conversational level. Structured machine feeds can rank higher. Only a true system of record produces top rank facts. Nothing inside a message can raise its own authority, so an email claiming to be from the CFO is still just an email, and it cannot overwrite a fact your system of record wrote.

A feed sends its key in a request header, never in the address, and every delivery is checked against it before anything is read.

Agents reading out#

AI tools and agents never touch a memory anonymously. Each one has its own identity and key, created on the Access and trust page, with a grant per room.

• Allow. Questions are answered at once, and the access is recorded.
• Approve. Questions wait for a named person to say yes in the Review queue before anything leaves the room. The decision and the decider become part of the record.
• Deny. Refused, and the refusal is recorded too.

Anything not named in the grant is denied by default. The technical details of asking, polling held questions, and connecting over MCP are in the Agent API.

When one agent asks another#

Agents never talk to each other directly. Agent A asks agent B's memory, using its own key, under the grant B's owner wrote. If the room requires approval, the question waits until B's owner decides. The refusals, the holds, and the answers all land in B's audit record under agent A's name.

Bringing in history#

Live sources only move forward. To bring months or years of existing conversation into a memory, exports are grouped into daily conversations, dated to their real timestamps, and read oldest first so later facts replace earlier ones the natural way. Contact us and we will help with your first import.

Watching it all#

The Access and trust page shows the agents and their grants, the connected feeds with their trust levels, and recent activity. The Review queue holds questions waiting for you, with one click approve and deny, and a badge in the navigation counts them.

Disconnecting a source#

Disconnecting deletes our copy of the credential at once. Where the provider allows it, our access ends there as well. Google Drive tokens are revoked at Google and the GitHub app is removed from the account that installed it. Microsoft and Discord do not offer a per connection revocation, so if you want to be certain, also remove Spark Memory from your Microsoft account's app list or from the Discord server. Removing a memory keeps its data unless you ask for it to be deleted, and deleted data is gone from our backups within fourteen days.