EveStudioDocs

Memory

An agent's usefulness depends on what it can remember. Eve gives every agent memory at two different timescales, and Studio's Memory tab shows you which one is active.

Two kinds of memory

Native durable sessions (built in). Every Eve conversation is a durable session. A session is the whole conversation or task: it is long-lived, survives process restarts and redeploys, and can span many messages over days or weeks without losing context. Within a session the agent keeps its full history — turns land in order, tool calls and their results keep their order, and you can read the whole thing back. When a session's context grows large, Eve compacts it: it writes a checkpoint that summarizes earlier history so the conversation can keep going without overflowing the model's context window. This is automatic and needs no configuration.

The boundary matters: durable sessions remember within a conversation. A brand-new session starts fresh. Subagents get their own separate session and context, so nothing crosses that boundary implicitly.

External long-term brain (across sessions). When an agent needs to remember things between sessions — facts, past work, what performed well, standing preferences — it uses an external memory store. Kybernesis's Arcana is the store these template agents use. Arcana is wired in as a connection that exposes memory tools (remember, recall, search) the agent calls during a turn, so memory is something the agent actively writes to and queries, not an ambient buffer. Each agent (and each subagent) can point at its own Arcana workspace, giving it a separate brain.

The two work together: durable sessions carry the current conversation, and the external brain carries what should outlive it.

The Memory tab

Open an agent and select the Memory top tab. It shows whether the agent is running on Eve's native durable sessions alone, or also has an external long-term brain (Arcana) attached — and, if Arcana is attached, the workspace it points at.

Wiring Arcana in Studio

To attach an external brain, provide three things in the Memory tab:

  • A workspace — the Arcana workspace (brain) this agent reads from and writes to. Different agents or subagents can use different workspaces to keep their memories separate.
  • An env var name — the environment variable the agent's connection reads its key from (for example ARCANA_API_KEY).
  • An API key — the Arcana key that authorizes access to that workspace.

Studio stores the key under the env var name you gave. If the agent is linked to Vercel, the key is also pushed to the project's environment variables, so the deployed agent has it too — not just your local dev server. (Linking is the same Connect to Vercel step the model already needs.)

Under the hood this configures a connection under agent/connections/ that points at Arcana and sends the key; the agent's own tools then surface and query that memory during a turn. Keep the key out of source — it lives as an environment variable, never in code.

Where to go next