Memories

The Memories page is the management surface for Auto Memory — the durable, long-term facts an AI agent decides are worth keeping and reuses in every future conversation. Where the Auto Memory tool lets an agent write memories at runtime, this page lets you read, correct, and prune them.

Agents create memory entries on their own; there is no manual "new memory" button on this page. You come here to audit what an agent has learned, fix a fact that drifted, retype a mis-classified entry, or delete something the agent should no longer act on.

Auto Memory vs. this page

Auto Memory complements chat memory: chat memory replays the current conversation, while Auto Memory persists facts across all future conversations of the same scope. See Chat memory vs Auto Memory for the distinction.

The name of each entry is set by the agent when it first stores the memory and is immutable on this page — everything else (title, description, type, and content) is editable.

Accessing Memories

Memories live under AI → Memories in the automation navigation, alongside Skills. The page has two regions in the left sidebar:

• The AI section nav — switch between Skills and Memories.
• The Type filter — narrow the table to a single memory type (see Filtering by type).

At the top of the page a header shows the active filter as the title, the total memory count, an environment selector, and a search box.

Selecting an environment

Memories are environment-scoped so that facts learned in one environment do not bleed into another (for example, DEVELOPMENT preferences do not surface in PRODUCTION). Use the environment selector in the header to switch environments; the table reloads to show only the memories stored for that environment.

Memory types

Every memory carries a type that classifies the kind of fact it holds. The type controls how entries are grouped in this UI and which slice of the index the agent sees at runtime.

Viewing the memories table

The table lists every memory in the current workspace and environment (optionally narrowed by the type filter and search). Each row shows:

When no memories exist for the current workspace and environment, the page shows an empty state explaining that the agent stores facts here as you chat with it — user preferences, project decisions, corrections, and external reference pointers — and that entries appear once the agent learns something worth keeping.

Filtering by type

Use the Type filter in the left sidebar to show only one kind of memory. The options are:

1. All (default) — show every memory.
2. User
3. Feedback
4. Project
5. Reference

The selected type becomes the page title, and the memory count updates to reflect the filtered set. The filter is applied on the server, so switching types reloads the table.

Searching

Type into the Search by title or description... box in the header to filter the visible rows. Search matches against both the title and the description of each memory and runs on top of the active type filter. Clearing the box restores the full (type-filtered) list. If the search matches nothing, the table shows "No memories match your search."

Viewing a memory

1. Click the View (eye) icon on a memory's row.
2. The detail dialog opens, showing:
   • The memory title and its type badge in the header.
   • The name and the last-updated timestamp beneath the title.
   • The description (if set).
   • The full content, rendered from Markdown.
3. Click Close to dismiss the dialog.

Editing a memory

1. Click the Edit (pencil) icon on a memory's row.

2. The Edit memory dialog opens, pre-filled with the memory's current values.

3. Change any of the editable fields:

   Field	Editable	Notes
   Name	No	Immutable here. To rename a memory, use the agent's rename tool. Shown disabled for reference.
   Title	Yes	Required — Save stays disabled while the title is empty. Up to 255 characters.
   Description	Yes	Optional one-line summary. Up to 1024 characters.
   Type	Yes	Choose USER, FEEDBACK, PROJECT, or REFERENCE.
   Content (Markdown)	Yes	The full memory body, edited as Markdown in a monospace text area. Up to 65,536 characters.

4. Click Save to persist your changes, or Cancel to discard them. If you make no changes, the dialog simply closes without a write.

On success a "Memory updated" confirmation appears and the table refreshes.

Deleting a memory

1. Click the Delete (trash) icon on a memory's row.
2. The Delete this memory permanently? confirmation dialog opens, warning that the entry will be removed from the agent's long-term memory.
3. Click Delete permanently to remove the memory, or Cancel to keep it.

On success a "Memory ... deleted" confirmation appears and the row disappears from the table.

Storage providers

By default, memories are stored in ByteChef's relational database. Self-hosted operators can select where they live with the BYTECHEF_AI_AUTO_MEMORY_PROVIDER environment variable (bytechef.ai.auto-memory.provider):

Selecting a provider whose file-storage backend is not configured fails fast at startup rather than falling back — a silent fallback would write memories to storage the operator did not choose. File backends are best-effort single-writer: concurrent edits of the same memory resolve last-write-wins, and there is no migration path between providers — switching does not move existing memories.

Frequently Asked Questions (FAQs)