I have notes from ten years ago that are still valuable.
If I ask:
**“What was I working on ten years ago today?”**
I want an AI system to find them.
If I ask:
**“What should I be working on today?”**
I probably don't want those same notes treated as current truth.
And if I ask:
**“How did my thinking about this problem evolve over the last ten years?”**
now I need both.
The information didn't become wrong because it got old, its relationship to the question changed.
That distinction is one of the reasons Phasoric exists, and it's the cleanest way I know to state the problem we are working to solve.
A repository that remembers everything is not necessarily intelligent
The industry's answer to context problems has mostly been *more*.
Bigger context windows, longer memories, retrieval-augmented generation over every document you own. Feed the model more information and trust it to sort things out, but more context is not the same as the right context.
A model with your entire history available does not automatically know which parts are still true, which facts supersede others, which sources are authoritative, or which information matters to the question being asked right now.
It just has more material to reason over, which can help in certain contexts and hurt in others. Ultimately, too much information can make answers worse.
A stale fact can outrank a current one and a correct answer can drown among plausible neighbors. Two contradictory statements can both be retrieved without any understanding of which one governs the present.
The hard problem is not memory capacity, but **context intelligence**.
If that sounds abstract, here's the ladder: the stages a context system has to climb, and what still breaks if you stop at each one.
Nine stages between storage and intelligence
1. Storage
This is where everyone starts because it's easy and feels like progress: save everything, worry about interpretation later.
Storage gets you nothing lost, because it is cheap. We can store everything, but it doesn't get you anything found.
2. Retrieval
Add search: keywords, semantic search, embeddings, and now information can be located, but retrieval finds things that *look like* answers.
Similarity is not relevance. The ten chunks most semantically similar to a question are not necessarily the ten pieces of context that matter most.
Retrieval gets you candidates, but it doesn't get you judgment or understanding of the right information.
3. Relevance
Now the system starts ranking information for this question, asked now, by this user, for this purpose.
This is the first rung that starts to feel intelligent, but relevance alone is still not enough because it can find the right fact for the topic without knowing whether that fact is current.
4. Temporal context
Facts have lifecycles.
Prices change. Teams reorganize. Plans get abandoned. Products get renamed. Policies are revised. Decisions are superseded.
A system that does not understand *when* a fact was true can confidently serve yesterday's truth as today's answer.
Temporal context lets the system distinguish:
**true then** from **true now**, but it still does not explain how facts relate.
5. Relationships
Facts do not live alone.
This decision blocked that project.
This requirement descended from that regulation.
This person owns that system.
This document superseded that one.
Without relationships, every answer is isolated.
Relationships tell you how information connects, though they do not tell you whether the information is trustworthy.
6. Provenance
Where did this come from?
Who recorded it?
When?
From which source?
Under what authority?
Provenance is the difference between:
**“The system says...”**
and:
**“This conclusion comes from these sources, recorded at these times, with this chain of evidence.”**
That makes an answer inspectable, auditable, and debuggable, but provenance alone does not resolve disagreement and real repositories contain disagreement constantly.
7. Contradiction and change detection
Two sources may disagree because one is wrong, and more often, they disagree because the world changed.
A useful context system has to recognize when statements conflict, understand whether one supersedes another, and preserve the reason for the change.
Sometimes that relationship can be inferred confidently from time, authority, or explicit revision history. Sometimes it can't and in those cases, the system should surface the contradiction rather than invent certainty.
Change detection gives you a coherent view of evolving truth.
8. Governed context
Who decides what is authoritative?
Who is allowed to change it?
Which sources govern which domains?
What requires review or approval?
What information can this user or agent access?
Governance can sound like bureaucracy until you need to trust the result.
It is the difference between a system you casually query and a system you are willing to rely on.
Governed context gives you answers you can defend, it still does not get anything done.
9. Action
At the top of the ladder is context strong enough to drive behavior.
Not simply:
**“Here's the answer.”**
But:
**“Here's the answer, here's why it matters to what you're doing, and here's the next action — with the authority, evidence, and guardrails necessary to take it.”**
Context that stops at answering is still primarily an information system, but context that can safely inform or execute action becomes operational intelligence.
Every rung depends on the ones below it.
Retrieval without storage is impossible. Relevance without retrieval has nothing to rank. Temporal reasoning without relevance becomes trivia.
Governance without provenance is theater. Action without governance is dangerous.
That is why context intelligence cannot simply be bolted onto an AI system after the fact, it has to exist in the architecture.
We didn't invent the example. We lived it.
A made-up example would be less useful, so here's a real one, ours.
Phasoric's history contains these two facts:
- **May/June 2026:** The product is called Basalt Notes. - **August 2026:** Basalt Notes was rebranded as Phasoric as the product evolved beyond notes into connected intelligence.
A naive retrieval system can return either statement.
Ask:
**“What is the product called?”**
and it might retrieve the June fact, or the August fact, or both.
A model handed both without sufficient temporal context might answer:
> “The product has been known as both Basalt Notes and Phasoric.”
Technically true, practically useless.
A context-intelligent system should understand both statements and the relationship between them.
**“What was the product called in June?”**
Basalt Notes.
The question is anchored in the past, so the June state is the relevant one.
**“What is the product called today?”**
Phasoric.
The later state supersedes the earlier one for a present-tense question.
**“How did the product evolve?”**
It began as Basalt Notes, with a stronger emphasis on notes and knowledge capture. As the product and architecture evolved toward connected intelligence, the old name no longer described the problem well enough, so the product became Phasoric.
Now the two facts do not compete, they work together. The second does not erase the first, it gives the first context.
Same storage. Same repository.
Completely different quality of answer, and this is not a hypothetical we constructed for a demo.
It is the product's own history.
If a context system cannot reason coherently about the product it runs in, nothing on the marketing page matters.
The bottleneck is moving
The next decade of AI progress will not come only from larger models, it will also come from better systems around those models.
The models keep improving, context windows keep growing, tool use keeps expanding, but the harder question is becoming:
**What should the model know right now?**
Not everything, not merely the most similar information, the right information.
Current where current matters. Historical where history matters. Connected to the right people, projects, sources, and decisions.
Grounded in provenance. Governed by authority. Aware of contradiction and change. Useful enough to support action.
The bottleneck is moving from model capability toward context quality, that is the problem Phasoric is built around.
I touched on a compressed version of this idea in my [first podcast appearance](https://youtu.be/JJotbDSicJA), where we discussed AI memory, context, and why more information is not automatically better information.
I wrote about the broader conversation and what changed in the month between recording and publication on [aketchel.com](https://www.aketchel.com/blog/ai-more-help-than-hype-real-technology).
If you're building a system where context is becoming the limiting factor, or your AI keeps being wrong for reasons that are not really model problems, that's exactly the class of problem we are interested in.
**[Explore Phasoric →](https://phasoric.com)**
Frequently asked questions
What is context intelligence?
Context intelligence is the ability to determine which information matters for a specific question or action based not only on similarity, but also on relevance, time, relationships, provenance, authority, contradiction, and change. The goal is not simply to retrieve information, but to provide the right information for the current situation.
Why isn't a larger AI context window enough?
A larger context window increases how much information a model can process, but it does not automatically tell the model which information is current, authoritative, relevant, or superseded. More available information can still produce poor answers when stale or contradictory context is treated as equally valid.
How is context intelligence different from RAG?
Retrieval-augmented generation, or RAG, retrieves information that may help a model answer a question. Context intelligence goes further by reasoning about which retrieved information is relevant, when it was true, how it relates to other information, where it came from, whether it has been superseded, and how it may safely be used.
Why does temporal context matter for AI memory?
Information changes over time. A statement may have been completely correct when recorded but no longer describe the current state. Temporal context allows an AI system to distinguish what was true then, what is true now, and how the state changed between them.
What role does provenance play in AI context?
Provenance records where information came from, when it was recorded, and the evidence or authority behind it. This makes AI answers more inspectable, auditable, and defensible instead of requiring users to accept an unsupported response from the model.
How does context intelligence help AI agents take action?
Agents need more than an answer before they can safely act. They need relevant and current information, appropriate authority, provenance, governance rules, and an understanding of uncertainty or contradictions. Context intelligence provides the foundation for deciding not only what an agent knows, but what it is permitted and justified to do.
What is the difference between AI memory and context intelligence?
AI memory focuses on preserving and retrieving information from previous interactions or stored knowledge. Context intelligence focuses on deciding which parts of that memory matter now. Memory answers “What do we know?” Context intelligence adds “What is relevant, current, connected, authoritative, and appropriate for this situation?”