Memory

Save what your workflow learns between runs, give it reference documents, and choose what each neuron recalls.

Memory is what a workflow knows and remembers between runs. It holds memories, which runs or you write and keep up to date, and documents, reference files you upload that runs read but never change. Use Memory when a workflow should build on earlier runs, or should always have a price list, a style guide or similar material at hand.

The Memory section

Open Memory in the canvas's left rail. The Memory section has three tabs: Memory, Review and Settings. The memory changes you made, with Undo, are listed under Versions in the left rail; see Undo your memory changes.

On the Memory tab:

Memory used shows how much of this workflow's memory space is in use, as a percentage. Hover over the info button to see the total across all your work.
Search memory… finds memories and documents by name or content.
MEMORIES lists what runs or you have written.
DOCUMENTS lists the reference files you uploaded.

Add a memory by hand

In the Memory section, click Add memory (below the lists). Keep Memory selected at the top of the form.
Type a Name, for example "Customer's plan".
Under What is this memory for?, say what belongs in it. The AI follows this description when it saves to the memory.
Optionally type the starting content under What should be remembered? You can leave it empty and let runs fill it.
Click Save.

Add a document to Memory

A document in Memory is reference material every run can read, such as a price list or a style guide. Runs never change it; only you can.

In the Memory section, click Add document (below the lists), or choose Document at the top of the Add memory form.
Choose Upload a file and pick the file, or choose From Documents to use a table, doc or slides from this workflow's Docs.
Check the Name.
Click Save. Reading and indexing a large file, a scan or audio can take up to a minute: keep the form open.

A document picked From Documents stays in sync with the Docs version. If the Docs version is deleted, the Memory copy shows Removed from Documents and keeps the last synced text.

Remember: save inputs and results to memory

Remember saves a neuron's input or result to memory after each run.

Open the neuron's GET (for an input) or MAKE (for the result).
Click Remember on the input or on the result.
Open the menu beside Remember to choose how it's saved: Automatic saves it right away; Propose sends each new value to Review first. Until you choose, it follows the workflow's default in Settings, which is Automatic for a new workflow.
Under SAVE TO, choose Choose automatically, This neuron's own memory, one of your memories, or New memory… (type a Memory name and click Create and use).

To stop saving, open the menu and choose Stop, or click Remember again. Only new details are saved; anything the memory already knows is not saved again.

A queue carries memory from one item to the next only when Remember is Automatic. Items saved as Propose wait in Review and are not read by later items. When that happens, the queue shows a note that you can dismiss, with a link to the Remember setting to change. Each queued item waits for the previous item's memory to be saved (up to a minute) before it starts, so the next item can use it.

A memory's settings

Click a memory in the MEMORIES list to open it:

WHAT THIS IS FOR: what belongs in this memory. Click the pencil to edit it and guide what the AI saves here.
SAVED IN MEMORY: what the memory holds right now. Click the pencil (Edit memory) to change it. Below it is the memory's reference, with a copy button (Copy ID). See References.
KEEP IT: how long the memory lasts: Auto: removed after 7 days without use, Until you delete it, For 7 days, For 30 days or For 90 days. Auto is the choice for a memory you add in a new workflow, or in an agent you install: it is removed once it has gone 7 days without being edited and without a run reading it, and a line says when. Every edit, every run that reads it and every Test this neuron run starts that time again. A memory that a neuron picks under Only items I pick, that a neuron uses as a {{memory.…}} reference (in its own text, or inside a workflow variable the neuron uses), or that is set to Ships with the agent is not removed while that is so, and its line says it is kept. A memory removed this way can't be restored. Workflows and installed agents that already existed keep the choice each memory already has. With a set number of days, START COUNTING sets whether the time counts From when it was first saved or From its last change, and a line says when it will be removed.
SOURCES: the runs and files the memory's content came from, once runs have written to it.
A line such as "Added by you · found by search in N neurons" (N is a count) says who wrote it and which neurons use it.

Don't use a memory

Don't use keeps a memory but stops runs using it. A memory set to Don't use shows NOT USED and is skipped by search and by references. Click Use again to turn it back on.

Delete a memory

Open the memory, click the bin icon (Delete), then Delete again to confirm. You can restore it with Undo in Versions for 7 days. After that it is deleted for good, with its content and files, and the space it used is freed. A reference to a deleted memory stops working until the memory is back.

A document's settings in Memory

Each document in the DOCUMENTS list has Open, Download, Replace and Remove. Removing a document picked from Docs takes it out of Memory only: it stays in Documents.

Review proposed changes

The Review tab lists what runs want to save when Remember is set to Propose. Runs never wait for you: the work carries on, and the proposal waits here.

For each proposal:

Save saves it to the memory.
Edit lets you change it before saving. Highlighted text is new.
Add as new memory saves it as a separate memory instead.
Discard throws it away.

If another run has changed the memory since, the card says so: click Reload suggestion to see the latest version.

Undo your memory changes

Open Versions in the canvas's left rail. Its History lists the memory changes you made, newest first, next to the workflow's versions, such as "You deleted …" or "You set … to Don't use". Click Undo on a change to reverse it, and Redo to apply it again, until something else changes. A change that was changed again later says "Changed since. Undo the later change first."

Undo is for your own edits only. Changes a run makes with Remember are not in the list and can't be undone from it. Memory still shows who wrote each item.
Changes are kept for 7 days, and at most the last 25 changes. After that they are removed from the list and can't be undone.
They count toward the workflow's storage (the same storage as Docs and dashboards), so a very full workflow can't keep new changes. The change is still saved; you are told it can't be undone. Deleting a memory is refused until there is room.
Ctrl+Z (Cmd+Z on a Mac) also undoes the memory changes you made in this browser tab, together with your canvas edits, in the order you made them. While you are typing in a text box, it first undoes the text you typed there; when the box has nothing left to undo, the next press undoes your last change. Older changes are undone from the list.
Restoring a version doesn't change memory. See Versions.

Memory settings

The Settings tab sets how the whole workflow remembers:

When something is remembered: the default for every Remember: Save automatically (saved right away; undo from Versions), the default for a new workflow, or Propose (waits in Review; runs don't pause). Each Remember can override it. Changing the default here doesn't change a Remember that already has its own choice.
Each neuron: what each neuron saves and reads. Click the gear beside a neuron to open its DO › Memory settings.
Three layers: what Users of your agent get. Builder knowledge is what you ship and is read-only for Users. User memory is what each User's copy learns, private to that User. Policy is how much a User's copy may learn.
Ships with the agent: turn on each memory that should be included when you publish the workflow as an agent.
What a User's copy may learn: Nothing (their copy only uses what you ship) or User decides (each User chooses Propose or Automatic and manages their memory on the agent page; see Using an agent).

What a neuron reads from memory

The Memory card in a neuron's DO tab sets what that neuron recalls each run.

Turn on Use memory in this run (on by default). When it is off, DO only sees its inputs; exact references still work.
Choose which memory it reads:
- Find what's relevant: Memory picks the items that fit each run. LAST RUN USED shows what the last run picked.
- Only items I pick: tick the items under ALWAYS INCLUDE. NEXT RUN GETS shows what the next run will read.

Saving to memory is set by Remember in GET and MAKE, not on this card.

Limits

Memory per workflow: up to 50 MB.
Memory across all your workflows and agents: up to 200 MB.
Memory per User's copy of an agent: up to 50 MB.
Memory name: up to 200 characters.
What this memory is for: up to 500 characters.
One memory's content: up to 256 KB.
Document upload: up to 32 MB per file; up to 8 MB of a document's text is kept.
Files Remember can read: up to 25 MB per file.
Undoable changes: the last 25 changes, for 7 days.
Find what's relevant: up to 3 items per run.

Common problems

Memory is full

The Memory is full notice means this workflow's memory has no room left, so the run's value wasn't remembered. The run itself finished. Click Open Memory, then free up space: delete memories or documents you no longer need, or shorten long memories. A proposal that couldn't be saved stays in Review until there is room.

Your memory is full

Your memory is full means your memory across all your workflows and agents has no room left. Free up space in any workflow's Memory, or in an agent setup's memory (its Settings tab on the agent's page; see Using an agent).

Not saved to memory

Not saved to memory means a neuron ran, but its Remember couldn't save to the memory it points at: the memory expired (it can't be restored: add it again in Memory), was deleted (if you deleted it in the last 7 days, restore it with Undo in Versions), or is set to Don't use. Click Open Memory. Turn a Don't use memory back on with Use again, or choose another memory under SAVE TO.

A memory disappeared on its own

A memory with a KEEP IT time is removed when that time runs out, and an Auto memory is removed after 7 days in which nothing edited it and no run (or test run) used it. An Auto memory can't be restored once it is removed. Set KEEP IT to Until you delete it on a memory that should stay. An expired memory can't be restored: add it again in Memory. A memory you deleted yourself can be restored with Undo in Versions for 7 days, and is deleted for good after that.

A run didn't use a memory I expected

With Find what's relevant, Memory picks only the items that fit each run. To make sure a neuron always reads a memory, choose Only items I pick and tick it under ALWAYS INCLUDE, or put a reference to it in the instructions. Memories set to Don't use are always skipped.