memories-guide

Memories inject user facts (preferences, projects, style) into the system prompt.

Workspace memories: Applied to chats in a single workspace. Limit: 20 per workspace.
Global memories: Applied to all workspaces. Limit: 5 total.
Access: Tied to user accounts. In multi-user mode, memories are private to each user.

Enabling Personalization

Open any workspace chat.
Click the settings icon (top right).
Select Memories from the menu.
Toggle Enable Personalization ON (Admin access required to toggle on/off; non-admins can manage their own memories once enabled).

Manage via the Memories sidebar:

Add: Click + in the tab header. Input a single-sentence fact.
Edit/Delete: Click the three-dot menu on the memory card.
Move: Use the three-dot menu to select Move to global or Move to workspace.

Uses the configured workspace chat model (requires tool-calling support) to extract facts from recent chats.

Runs only when:

Observer: Identifies up to 3 candidate facts per run with confidence ratings.
Reflector: Compares candidates against existing memories to classify scope (global vs. workspace), deduplicate, consolidate updates, and filter low-confidence items.

When Personalization is active, AnythingLLM appends the following to the system prompt:

All global memories (max 5).
Top 5 workspace memories (selected by relevance to current chat history; falls back to the 5 most recent if relevance scoring fails).

Note: Memory content is sent to the LLM provider. Do not store credentials or sensitive data.

Define these variables in the instance .env file:

Variable	Default	Description
`MEMORY_EXTRACTION_INTERVAL`	`3hr`	Background extraction job run frequency.
`MEMORY_IDLE_THRESHOLD_MS`	`1200000`	Minimum idle time (ms) required before processing a workspace (Default: 20 mins).