Agent memory

Large Language Models (LLM) are strong at conversation. In A1KnowHow you turn them into agents by giving them tools — search your documents, open a file, add a line, add external integrations and so on. The catch is that each new chat starts fresh: the agent does not automatically remember what you said last week, and you end up repeating the same details (names, dates, preferences) again and again.

The memory pattern fixes that in a simple way. You keep a normal document—often called MEMORY.md—and you tell the agent in its instructions to read it before answering and add facts to it after. Whatever lands in that document is still there the next time you chat, or when another agent uses the same instructions.

There is no separate “memory” product in the app. Memory is just a document in a workspace that you choose on purpose. Because it is an ordinary file, you stay in control: if a fact changes, or the agent adds something wrong (for example a typo in a client’s name), open the document in the editor and fix it—the same way you would edit any other note.

How it works

1. You create MEMORY.md in a workspace (same steps as Create Document; the filename is up to you, but MEMORY.md is a common convention).
2. You tell the agent (in Settings → Agents → your agent → System Prompt) which document is memory and what to do: read it before answering, add lines after when something should stick. See Creating agents.
3. The agent reads the document at the start of a turn and uses what is already written.
4. The agent answers the user.
5. The agent adds short lines to the document when a fact, preference, or date is worth keeping.
6. Next session—tomorrow, next week, or a different chat—the agent reads the same document again. The facts are still there.

Any other agent whose instructions point at the same document can read and update those same facts. For example, a chat agent and a workflow step can share one memory document.

Memory is a normal document

MEMORY.md is not built into A1KnowHow. You choose the name and the workspace. You can open it in the editor, change lines yourself, or share the workspace with teammates so everyone sees what the agent recorded.

That also means memory is permanent across sessions in the straightforward sense: whatever is written in the document stays until someone changes or deletes it—and you can make those changes anytime.

Create a workspace for memory (in the app)

1. In the left sidebar, click Organize.
2. On Organize Knowledge, click Create workspace (top right).
3. Enter a name, e.g. Skills & Memory – UK Remortgage, and save.
4. Click the new workspace card to open it.

You can keep skills in the same workspace or add a child workspace later: on the workspace page, open the actions menu (chevron on the workspace card or toolbar) and choose Add Workspace for a nested folder (e.g. Skills only).

Create your memory document

Inside the workspace you chose for memory:

1. On the workspace page, under Documents, click Create Document (toolbar button), or open the workspace actions menu and choose Create Document.
2. Name the document MEMORY.md.
3. Paste a starter template:

# Memory — Remortgage Assistant (UK)

Curated facts, preferences, decisions, dates, and lessons for this user.
Most recent entries at the bottom. Each line looks like:

- [YYYY-MM-DD] <scope>: <content>

Scopes: pref | decision | fact | lesson | date | correction

## Entries

<!-- The agent appends below this line. -->

Save the document. Copy the document id and workspace id from the browser URL when the file is open (see Building your agent — Copy the document and workspace IDs). The agent needs those ids in its instructions—not just the filename MEMORY.md.

Tell the agent to use memory

1. Settings (sidebar) → Agents → open your agent, or Add New Agent / From Template → Agent - With Memory (Creating agents).
2. In System Prompt, add instructions along these lines:

• Before you answer, open the memory document and read the entries (for example the first few hundred lines).
• Treat what you read as facts about this user unless they say otherwise.
• Prefer recent entries if two lines conflict.
• After you answer, if something durable should be remembered, add one to three lines to the memory document.
• Only add to that document—do not change other documents unless the user asks.

When you use a workflow, the template Agent - With Memory fills in similar instructions and asks you to pick the memory workspace and document.

In the System Prompt, include the memory document id (and workspace id if helpful) so the agent opens the correct file to read and append. Workflow templates fill <<memory_doc_id>> for you; standalone agents need the id pasted from the URL. What matters is the behaviour: read first, update after, only on that document id.

What the agent should remember

Worth writing down	Skip writing down
Preferences (“prefers 5-year fixes”)	Small talk and pleasantries
Decisions (“chose Nationwide tracker on 2026-04-15”)	Live rates or product codes that go stale quickly
Key dates (“current fix ends 2026-11-30”)	Things the user asked you to forget
Corrections (“this is a remortgage, not a first purchase”)	One-off clarifications that will not matter later
Lessons (“lender X only ports if completion is within 6 months”)

Use one line per entry, for example:

- [2026-04-21] date: current fix ends 2026-11-30
- [2026-04-21] pref: wants to avoid fees added to the loan

Example: before and after one chat

Before (empty memory):

## Entries

<!-- The agent appends below this line. -->

The user asks: I’m 8 months from the end of my 2-year fix. Help me plan.

The agent answers with general timing advice, then adds to the document:

## Entries

- [2026-04-21] date: current fix ends 2026-11-30 (user stated, 8 months from 2-year fix)
- [2026-04-21] fact: on a 2-year fixed rate, remortgage planning

Next week the user asks: When should I start shopping?

The agent reads those lines first, so it does not ask again for the fix end date. It can say something like: “Based on your fix ending in November 2026, starting to compare deals around May 2026 is reasonable…”

Use memory in a workflow

In Workflows → Create workflow, add a step from template Agent - With Memory. Choose your agent, then fill in:

• Memory workspace — where MEMORY.md lives
• Memory document — your MEMORY.md

When the workflow runs, the step uses the same read-and-update behaviour. Facts written in chat are visible in the workflow, and the other way around, if both use the same memory document.

Next step

Memory gives your agent a notebook. Agent skills add reusable playbooks as documents in another workspace.