eve Memory Agent

The eve Memory Agent is a Slack-based AI assistant template built on the eve agent framework that stores the preferences people share with it and recalls them at the start of every conversation. It runs on Vercel and ships as a cloneable repository within Vercel's Agent Stack, so deploying it creates a Git repository, a Vercel project, a Slack app installed through Vercel Connect, and a private Vercel Blob store for memory.

What is the eve Memory Agent?

The eve Memory Agent is a deployable Slack assistant template maintained under the vercel-labs organization that takes Slack messages as input and replies in-thread while persisting two kinds of remembered preferences: one set per Slack user and one set per channel. It is built with the eve agent framework, integrated with Slack via Vercel Connect, stores memory in Vercel Blob, and executes tasks in Vercel Sandbox. The repository is MIT licensed.

What makes it stand out?

• Two separate memory scopes — user preferences are shared by one Slack user across every channel and DM, while channel preferences are shared by everyone in one channel; each is defined as a file under agent/memory/.
• Recall before every reply — eve recalls both memories before responding to each message, and the model saves or removes entries through each memory's own tools.
• Identity-tied memory — memory is keyed to the verified Slack user and channel, never to text the model or user writes; DMs have no shared channel, so they use user memory only.
• Slack profile context — the agent looks up Slack profiles (names and profile images, never email addresses) for everyone in a thread, including mentioned people, with SLACK_IDENTITY_EXCLUDED_USER_IDS to skip bots.
• Preference precedence rules — agent/instructions.md tells the model to use channel memory only when someone states a channel-wide preference, and to follow a user's own preference over the channel's.
• Local dev parity — pnpm dev opens eve's interactive chat in the terminal using just-bash instead of Vercel Sandbox, with in-memory storage cleared on restart.
• Validation and deploy pipeline — pnpm validate, pnpm exec eve info, and pnpm build check changes before vercel deploy --prod.
• Vercel Drives extension — longer team-maintained documents such as runbooks can be mounted read-only into the sandbox alongside conversational memory.

Who is it for?

• Teams adopting an internal Slack assistant: deploy the template to get a working agent that remembers how individual colleagues like answers formatted, without building memory infrastructure.
• Slack workspace admins: set channel-wide conventions once that the agent applies for everyone in that channel, with per-user preferences overriding them.
• Agent developers: use the memory file structure (description and scope fields) as a starting point for custom memory providers described in CUSTOMIZATION.md.

How does it work?

Deploying through the Vercel button creates a repository and project, then you install a Slack app via Vercel Connect and connect a private Blob store. After deployment, add the agent to a channel and mention it to start a conversation; preferences you share are stored per the memory you state them in and recalled in later sessions.

FAQ

Does the eve Memory Agent work in direct messages?

Yes, but DMs only use user memory, because a direct message has no shared channel for channel-level preferences to attach to.

Where is memory stored during local development?

Locally with pnpm dev, memory lasts only while the development server runs and is cleared on each restart, and every conversation shares one user memory and one channel memory.

Can the agent share long reference documents?

Memory suits short preferences; for longer files like runbooks or internal glossaries, mount a Vercel Drive into the sandbox so the agent can search but not modify them.