Agent Memory

Pro+Agent Available on the Pro+Agent tier.

The agent remembers things between conversations. Not just what you said five minutes ago — facts about your world, preferences you’ve expressed, patterns it has noticed. Memory is how the agent becomes useful over time rather than starting from scratch every session.

Memory is separate from your workspace content. Your pages, tags, and links are the primary knowledge base. Memory supplements that with observations, preferences, and contextual knowledge that don’t belong in a page.

How memory works

Behind the conversation interface, the agent operates as a team. The Orchestrator — the agent you talk to — observes your conversation and extracts facts worth remembering. A dedicated specialist called the Archivist decides what to keep, where to store it, and when to surface it again.

At the start of each conversation, the Archivist loads relevant memories and provides them as context. During the conversation, the Orchestrator passes observations to the Archivist for storage. When the conversation ends, the Archivist reviews a summary and extracts anything it missed.

Memory scopes

Memories are stored at four scopes, from narrowest to broadest:

Scope	What it holds	Lifespan
Conversation	Working context for the current session	Persists across the conversation lifecycle
Channel	Topic-scoped knowledge across related conversations	Persists across conversations in the same channel
Workspace	Project facts and knowledge	Persists indefinitely
Account	Cross-workspace preferences and patterns	Persists across all workspaces

Workspace memories are scoped to the current workspace only — “The villain’s name is Malachar” stays in your fantasy novel workspace. Account memories travel with you — “I prefer concise responses” applies everywhere.

Memory lifetimes

Within each scope, memories have a lifetime:

Lifetime	Behavior
Long-term	Persists indefinitely. Survives consolidation and restarts.
Short-term	Persists days to weeks. Subject to gradual relevance decay.
Conversation	Cleared when the conversation ends. Working memory only.

Relevance decay

Short-term memories fade over time. Memories at different scopes decay at different rates — channel memories decay faster than workspace memories, which decay faster than account memories. This reflects how topical context becomes stale faster than project knowledge.

Frequently accessed memories resist decay. If the agent keeps retrieving a memory because it’s relevant to your work, it stays strong. Important memories that go unused gradually fade and are eventually pruned during consolidation.

Channels

Channels are topical rooms that group related conversations. When you have multiple conversations about different aspects of your project — one about character arcs, another about world geography — the Archivist can group them into channels so memories from each topic stay organized.

Channels are system-managed: the Archivist creates and assigns them based on conversation topics. Every workspace has a default channel for conversations that don’t fit a specific topic.

Conversations

Each agent interaction is a conversation — a named session within a channel that tracks the full lifecycle of the interaction. Conversations move through three states:

Status	What it means
Active	The agent is currently working in this conversation.
Idle	The agent session has ended. The conversation is available for review.
Archived	The conversation has been archived. Excluded from default views.

Conversations are grouped into channels — topic-based rooms that keep related interactions together. Every workspace has a default “general” channel.

Working Notes (scratchpad)

Each conversation has a scratchpad — a per-session notepad where the agent stores working state as key-value pairs. Unlike memories, the scratchpad is not included in every message. The agent reads it on demand when it needs to recall intermediate results or context from earlier in the conversation.

Working Notes appear in the Memory view under the Conversation scope alongside learned memories. They display with a Working Note badge to distinguish them from memories.

The scratchpad content is preserved when a conversation is archived, so you can review what the agent was working on. Working Notes are cleared by the agent at the start of a new session, not on archive.

Store and retrieve memories

You don’t need to manage memory explicitly — the agent decides when to store and retrieve facts based on context. You can also direct it with natural language:

• “Remember that Magistrate Thorne is secretly allied with the rebels.” — stores as workspace-scoped
• “Across all my projects, I prefer third-person limited POV.” — stores as account-scoped
• “For this conversation, we’re focusing on The Shattered Spire plotline.” — stores as conversation-scoped
• “What do you know about the politics of the Northern Provinces?” — triggers memory retrieval

The agent combines memory retrieval with workspace search — it checks both its memories and your actual pages to give the most complete answer.

Manage your memories

The Memory page provides a full view of everything the agent remembers. From here you can:

• Browse memories across all scopes
• Search by keyword or topic
• Edit existing memories to correct or refine them
• Delete memories that are no longer relevant
• Filter by source to see which agent type produced each memory

See Also

• Advanced Agent Features — Scheduled activities, skills, memory, and presence
• Agent Conversation — The conversation thread
• Getting Started with Agent — Start the agent