I've managed hundreds of people across four startups. One of the most frustrating things about hiring someone new? Repeating context. Explaining the same backstory, the same priorities, the same quirks of how we work — every single time.
Then I got my first AI worker. And somehow, the problem was worse.
The amnesia problem nobody talks about
Most AI tools today have the memory of a goldfish on a bad day. You spend 20 minutes explaining your project, your preferences, your constraints. You close the tab. You come back tomorrow. And it greets you like a stranger.
This isn't a minor inconvenience. It's a fundamental failure that makes AI feel like a toy instead of a teammate — the gap we keep coming back to in From chatbot to coworker.
Here's what breaks down when your AI has no memory:
- You repeat yourself constantly. Every session starts with "remember when I told you…" — except it doesn't remember.
- Quality degrades over time. Without memory, your AI can't learn what works and what doesn't. It makes the same bad suggestions on day 90 that it made on day 1.
- Context gets lost at the worst moments. A follow-up from last week's conversation? Gone. That decision you explained your reasoning on? Evaporated.
- You lose trust. When a tool forgets what matters to you, you stop relying on it for anything important.
According to McKinsey's 2024 State of AI report, 72% of organizations now use AI in at least one business function. But most implementations still suffer from this same flaw: they forget everything between sessions.
Let's say you're an operations lead who spent Monday walking a chatbot through your escalation policy so it could draft vendor replies. On Thursday, a vendor pushes back on terms — and the tool has no idea the policy exists. You either re-explain it or fix the draft yourself. Either way, you just became the memory.
Why context windows aren't memory
There's a common misconception that bigger context windows solve the memory problem. They don't.
A 128K token context window gives you maybe 100-150 turns of conversation before early context gets buried. And buried context might as well be deleted — the model effectively ignores it. More tokens doesn't mean better memory. It means more noise.
| Bigger context window | Real memory | |
|---|---|---|
| Scope | One conversation | Every conversation, across weeks and months |
| What happens to old context | Gets buried and effectively ignored | Stored, then retrieved when relevant |
| Selection | Everything in the window, relevant or not | The right context at the right moment |
| Effect over time | More tokens, more noise | The worker gets better at its job |
Real memory isn't about stuffing more text into a single conversation. It's about knowing what matters across conversations, retrieving it at the right moment, and connecting dots between things you said last Tuesday and what you need today.
What real AI memory looks like
True memory for AI workers needs three things:
- Persistence across sessions. When you tell your AI worker something on Monday, it should still know it on Friday. And next month. Without you repeating it.
- Intelligent retrieval. Not everything from every past conversation is relevant right now. Good memory means surfacing the right context at the right time — not dumping your entire history into every response.
- Learning from patterns. Over time, an AI worker with real memory should get better at its job. It should learn your communication style, your decision patterns, your priorities — the same way a human colleague does after working with you for months.
This is the difference between a chatbot and a coworker. A chatbot answers questions. A coworker remembers your last conversation, knows your preferences, and gets better every week. It's also a core part of what separates AI workers from AI agents: an agent executes and resets, while a worker accumulates context inside a role. (If the terms are new, start with what AI workers are.)
What we built (and why it took so long)
At Spinnable, we just shipped Conversation Memory — and honestly, it took longer than I'd like to admit. The reason: most "memory" solutions are just fancy search over chat logs. That's not good enough.
Our approach works differently. Before responding, your AI worker automatically retrieves relevant context from past conversations. Not all context — relevant context. It pulls in prior exchanges based on what you're discussing right now, so it has the background it needs without drowning in irrelevant history.
The result: responses that feel like talking to someone who actually remembers what you've discussed. Not perfectly — we're honest about that, and if you need guaranteed verbatim recall of every detail ever mentioned, no AI memory system today delivers that. But meaningfully. The kind of memory that makes you stop re-explaining yourself and start trusting the tool with real work.
The practical difference
Here's what changes when your AI worker actually remembers:
- Week 1: You explain your communication preferences, introduce your team, describe your workflows — the same investment you'd make onboarding any new team member.
- Week 2: You ask it to draft an email. It already knows your tone, who your key contacts are, and what projects are active.
- Week 4: It flags something you mentioned three weeks ago that's now relevant to today's conversation. You didn't ask it to connect those dots — it just did.
This is what turns an AI tool into an AI worker. Not more features. Not more integrations. Memory. The same thing that makes any human colleague actually useful after their first month.
It's also why memory compounds fastest in persistent roles. An AI executive assistant that remembers who your key contacts are, or an operations manager worker that remembers why last quarter's process changed, gets more valuable every week — because the role gives the memory something to accumulate around. As we argue in AI doesn't need better prompts — it needs a job description, context belongs to the role, not to the prompt.
What's next
Memory is just the beginning. Once an AI worker can remember, it can learn. Once it can learn, it can anticipate. We're building toward AI workers that don't just respond to what you ask — they proactively surface things you need before you realize you need them.
But that starts with solving the amnesia problem. And for the first time, we think we have.
Frequently asked questions
Why do most AI tools forget everything between sessions?
Because they rely on a single conversation's context window rather than persistent storage. When the session ends — or when the conversation grows long enough that early context gets buried — that information is effectively gone, and the next session starts from zero.
Don't bigger context windows solve the memory problem?
No. A 128K token window covers maybe 100-150 turns before early context gets buried, and buried context might as well be deleted — the model effectively ignores it. More tokens means more noise, not better memory.
What does real AI memory require?
Three things: persistence across sessions (Monday's context still exists on Friday), intelligent retrieval (surfacing only the relevant context at the right moment), and learning from patterns (getting better at the job over time, like a human colleague after a few months).
How does Spinnable's Conversation Memory work?
Before responding, the worker automatically retrieves relevant context from past conversations — not the full history, just the exchanges related to what you're discussing now. That gives it the background it needs without drowning in irrelevant detail.
Is AI memory perfect?
No, and we're upfront about that. Conversation Memory makes recall meaningful rather than perfect — the goal is that you stop re-explaining yourself and start trusting the worker with real work, not that it reproduces every detail verbatim.
How quickly does memory change the experience?
Within weeks. In week 1 you explain preferences and workflows; by week 2 the worker drafts in your tone and knows your contacts; by week 4 it connects things you said earlier to what's relevant today — without being asked.
Stop being your AI's memory
If you're re-explaining your business to a tool every morning, you're doing the remembering for it — and that's the job the AI was supposed to take off your plate. Hire a worker that accumulates context instead.
Create your first AI worker with Spinnable — it takes about 10 minutes to hire a worker that actually remembers.
