ArticlesWhere AI is going
Where is AI going, and where does memory fit?
Our last two articles were about memory for coding agents: why your engineering team needs one and how to choose one. This one steps back to the bigger picture.
AI is quickly getting good enough to think and act on its own. Three shifts are happening at once. Models now reason close to a human level. AI is starting to reach out instead of waiting to be asked. And AI is starting to work like a team member rather than a tool.
Together, they move the bottleneck. What holds AI back next isn’t intelligence. It’s knowing what you’re trying to achieve.
Is human-level reasoning arriving?
On the hardest reasoning tests, AI went from single digits to most of the way there in under two years.
In early 2025, the hardest reasoning tests were still out of reach. When Humanity’s Last Exam came out, with 2,500 questions written by nearly 1,000 subject experts to sit at the edge of human knowledge, the best models scored in the single digits. Today the top model scores 74% on Artificial Analysis’s independent evaluation.
ARC-AGI tells the same story. ARC-AGI-2 was built from puzzles people solve easily and AI found hard. It’s now largely solved, with top scores on the ARC Prize leaderboard around 95%. Its successor, ARC-AGI-3, tests whether AI can work through new games as efficiently as a person, and the best score has already passed 80%.
Benchmarks aren’t the job, but the direction is clear. Models can now work through an unfamiliar problem the way a strong professional does: break it down, try approaches, check the work.
The catch is one every manager knows. A brilliant new hire on day one is still new. They don’t know your goals, what’s already been decided, or why. Intelligence gets them through the problem in front of them. It doesn’t tell them which problem matters.
Is AI starting to reach out first?
Yes. AI is starting to speak up instead of waiting to be asked.
Until recently, AI waited for you. You asked, and it answered. That’s changing.
In September 2026, Meta launched Muse, a personal AI agent that, in Meta’s words, “can make suggestions unprompted” and “comes back when something changes or when it needs approval.” It can warn you before a registration deadline closes or book the dinner you mentioned, with controls to make it more or less proactive.
At work, the same shift is underway. Coding agents now run for hours in the background and come back with results, questions or a decision to approve.
Reaching out is only useful at the right moment, about the right thing. To get that right, AI has to know what you’re trying to achieve and what’s already settled. A reminder about a goal you dropped last month is noise. An agent that reaches further than it should is overreach: within weeks of Muse’s launch, some users reported it trying to read emails and private messages without permission. Proactive AI needs judgment about what matters, and judgment needs context.
What comes after copilots?
AI employees: AI that works like a team member, not a tool you prompt.
Think of what you’d expect from a good hire:
- They understand what the manager is after, not just the task.
- They work out the plan with the manager.
- They commit to weekly deliverables and report back.
Technical roles will come first: engineers, scientists and analysts, because the work is already digital and the results can be checked.
What separates an employee from a tool isn’t any single answer. It’s judgment over weeks: remembering what was agreed and why, knowing when a change matters enough to raise, and not redoing what a teammate already settled.
That judgment depends on memory of the strategy, not a log of every conversation. Your best colleague doesn’t remember every meeting word for word. They remember what you’re trying to do and where things stand.
What do these have in common?
Intelligence is no longer the bottleneck. Context is.
Better reasoning, proactive AI and AI employees all run into the same wall. The model is capable enough. What it lacks is the big picture.
A leader expects a team member to understand the big picture, and not to be thrown by details. A small change shouldn’t derail the plan. A real change of direction shouldn’t slip by unnoticed. That isn’t about intelligence: a very smart person who just walked into the room will still get it wrong.
More facts don’t fix this. Most AI memory today keeps facts, notes and documents, and finds the ones closest to the request. That helps, but a pile of facts can’t tell you which goal a fact serves, which question it answers, or whether it still matters. Give a capable model more facts and you get a more confident guess, not better judgment.
Why is strategic memory the key?
It gives AI the big picture: your goals, the questions under them, and the decisions with their reasons.
We think the missing piece is strategic memory: memory organized the way a team actually works through a problem. It keeps:
- Goals: what you’re trying to achieve.
- Key questions: what has to be figured out to get there.
- Decisions: the answer to each question, with the reasons and the options ruled out.
That structure is what makes judgment possible:
- It knows where things are going. A new session, or a new teammate, starts from the goals and where the project stands, not from the latest fact.
- It can tell a detail from a change of direction. A small change updates one answer. A new constraint that touches a goal reopens the questions under it.
- It asks the right questions. It organizes the work into the questions that matter and what’s been decided, the way a good lead keeps track of a project. Weeks of work become a handful of goals and questions, so AI gets more understanding from less.
And it’s what each of the three shifts needs:
- Proactive AI needs it to know when to speak up, and when to stay quiet.
- AI employees need it to keep commitments and judgment across weeks.
- Teams of people and AI need it shared, so everyone works from the same plan.
What does this mean for teams today?
It’s already happening in software.
Coding agents are the first place AI works this way every day: long sessions, real decisions, many agents on one codebase. Teams using them already feel the cost of agents that don’t know the plan: re-explaining, repeated experiments, a teammate’s decision quietly undone.
So treat memory as a strategic asset, not a feature:
- Own it. Your team’s goals and decisions are some of the most valuable things it produces. Know where they’re kept, and make sure you can take them with you.
- Share it with clear boundaries. The team should work from one picture. A contractor should see their project, not everything.
- Keep it current. Memory that still holds last month’s plan is worse than none.
And ask three questions about any AI you adopt:
- Does it know what we’re trying to achieve, or only what we told it last?
- Can it tell a small detail from a change of direction?
- Who owns what it remembers, and who can see it?
Where does that leave us?
The smartest AI will still need to know your plan.
AI will keep getting smarter. That’s the easy part to predict. The harder part, and the more valuable one, is giving it the big picture: what you’re trying to achieve, what’s been decided, and why.
That’s what we build. Assertion is strategic memory for your team’s coding agents today, in Claude Code, Cursor and Codex, or in our own agent. It’s free to start: try it on a real project.
Questions people ask
Is AI reasoning at human level yet?
On some of the hardest tests, it’s close. Top models score 74% on Humanity’s Last Exam, up from single digits in early 2025, and around 95% on ARC-AGI-2. Reasoning well isn’t the same as knowing your goals, though, and that’s the gap memory fills.
What is proactive AI?
AI that reaches out without being asked: a reminder before a deadline, a suggestion when something changes, a question when it needs approval. Meta’s Muse, launched in September 2026, is one example.
What is strategic memory?
Memory organized around what a team is trying to achieve. It keeps the goals, the key questions under each goal, and the decisions that answer them, with their reasons. So it knows where the work is going and can tell a detail from a change of direction.
Sources
- Humanity’s Last Exam, Center for AI Safety and Scale AI
- Humanity’s Last Exam results, Artificial Analysis
- ARC Prize leaderboard, with ARC-AGI-2 and ARC-AGI-3
- Introducing Muse, Meta, September 8, 2026
- Meta introduces Muse as a proactive personal agent, TestingCatalog
- What is Meta Muse?, Built In
Last checked November 17, 2026. Assertion gives Claude Code, Cursor and Codex strategic memory that writes itself as your team works. It is free to start: install it here.