Memory — how your agent learns you

Your agent has persistent memory — it remembers you from conversation to conversation. This is the single biggest thing that makes it different from a chatbot. Here’s exactly how it works.

Two memory files

Your agent maintains two small, structured files:

USER.md — what your agent knows about YOU

Seeded from your onboarding answers (name, role, company, goals, challenges, tools, timezone). Your agent updates it over time as it learns more about you — things you mention in passing, preferences you express, working patterns it observes.

Example entries:

“User prefers direct communication, no corporate fluff”
“User runs a solo consulting practice for small SaaS companies”
“User is most productive in the morning and prefers to schedule meetings after lunch”
“User’s main time drain is inbox triage”

MEMORY.md — environment facts and lessons

Things about your work, your tools, your world, and lessons learned from past tasks. Not about you personally — about the context you operate in.

Example entries:

“Top competitors are Acme Corp and Rival Inc. Both update their pricing page monthly.”
“When drafting LinkedIn posts, user prefers no emojis and no hashtags.”
“User’s morning briefing should be delivered at 7:15am every weekday.”
“User’s Acme Corp contact is Sarah Chen (sarah@acme.com), VP of Product.”

How memory gets injected

At the start of every conversation, both files are loaded into your agent’s working context. It can see all of your USER.md and all of your MEMORY.md before it generates a single word. That’s why it can say things like “Good morning, Maxson — I checked the Acme Corp site overnight and noticed they raised prices” without you reminding it who Acme Corp is.

How memory gets updated

Every 10 conversation turns, your agent spawns a background review — a disposable copy of itself that looks at the conversation and asks: “Is there anything here worth remembering long-term?”

If there is, it writes a new entry to MEMORY.md or updates an existing one. This is how the learning happens. You don’t have to tell your agent “remember this” — it decides for itself what’s worth keeping.

You can also force a memory update by saying things like:

“Remember that I prefer morning meetings”
“Make a note that our Q2 revenue was $47k”
“Don’t forget — the Acme deal is on hold until September”

It’ll add those to MEMORY.md immediately.

Size limits (and self-pruning)

Memory files have hard caps:

MEMORY.md: ~2,200 characters (roughly 400 words)
USER.md: ~1,375 characters (roughly 250 words)

Why so small? Because every entry costs tokens on every conversation, and big memory files get expensive fast. Your agent is disciplined about what it keeps.

When MEMORY.md gets full, your agent prunes the oldest / least important entries automatically to make room for new ones. You can also prune manually from the Memory tab in the dashboard.

Viewing and editing memory

From the dashboard, click Memory in the sidebar. You’ll see both files in plain markdown. You can:

Read everything your agent currently knows
Delete entries you don’t want anymore
Add new entries yourself
Reset either file entirely

Memory is YOUR data. It stays on your own agent’s persistent storage. It never gets shared with other users. It never gets used to train models.

Skills — the other kind of memory

There’s a third kind of memory: skills. When your agent figures out a non-trivial way to do something (a multi-step workflow, a tricky research approach, a careful email tone), it can write that as a reusable “skill” — a markdown document the next similar task can reference.

Skills live in the skills/ directory of your agent’s workspace. You can see them from the dashboard’s Workspace tab. Unlike MEMORY.md, there’s no hard size limit on skills — your agent can build a whole library of them over time, each scoped to a specific type of task.

What memory does NOT capture

The full text of every conversation. Past conversation sessions are stored separately in a searchable database (FTS5). When your agent needs to remember a specific past exchange, it does a search on the session database rather than loading everything into memory.
Things in your inbox, calendar, or files. Those stay in their own systems. Your agent reads them on demand; it doesn’t “remember” them in the MEMORY.md sense.
Anything you asked it to forget. Delete an entry and it’s gone.

Privacy

Memory lives in your private container, encrypted at rest
Nobody on our team can read it without your explicit permission
If you cancel your subscription, your memory files are retained for 30 days and then permanently deleted along with your container
You can export your MEMORY.md and USER.md at any time from the dashboard