ChatGPT, Claude, Gemini—try asking the AI you use every day this question.
"Do you remember the rules we decided on last week?"
The answer, nine times out of ten, will be: "Sorry, I don't have access to the content of our previous conversation."
The conclusion you reached after two hours of discussion with an AI. The design philosophy you refined through countless revisions. The workflow you promised to follow next time. The moment the session switches, all of it disappears.
Every time, it's like meeting for the first time. Every time, introductions from scratch. Every time, the same explanations repeated.
I use LLMs daily. Music production, programming, writing, design discussions—AI is no longer an "occasional tool" but part of my thinking process every day.
Which is why this amnesia was fatal.
The musical concept I spent two hours developing yesterday. The style direction I refined through repeated feedback. The design decision I made: "Let's use this approach next time." I open a new session the next morning, and the AI knows none of it. I have to explain everything from the beginning again.
If you're just using LLMs as a "convenient tool," this might be a minor inconvenience. But the moment you try to use AI as a daily partner—a work collaborator, a creative accomplice, a thinking partner—this "amnesia" becomes fatal.
If we're going to call them "partners," I want to continue from where we left off yesterday. But somehow, that's impossible. This was the starting point of this project.
Why Does AI "Forget"?
Let's develop a rough understanding of how LLM (Large Language Model) memory works. We'll cover technical details in later parts, but for now, just the intuitive picture.
Context Window: AI's "Short-term Memory"
LLMs have a concept called the context window. Roughly speaking, it's the amount of text an AI can see at once.
As of 2024, Claude 3 has about 200,000 tokens (roughly 100,000 characters in Japanese), and Gemini 1.5 Pro exceeds 1 million tokens. By 2026, this has expanded further. It might seem sufficient at first glance.
But this window includes the system prompt (the AI's configuration and instructions). It includes external tool definitions. It includes past conversation history. The area actually available for user conversation is much narrower than expected.
And most critically, information beyond the window physically ceases to exist. If the 100,000-character window receives 110,000 characters, the oldest information gets pushed out and disappears.
Compaction: Forgetting Disguised as "Summary"
Many AI agent frameworks handle this problem through compaction (compression). Old conversations are summarized and compressed into shorter text that stays within the window.
It seems reasonable on the surface. But summarization has a fundamental problem.
Information degrades the moment it's summarized.
"On March 5th at the meeting, Ramo decided to change Project A's design approach from flat UI to Material Design. The reasons were mobile tap area visibility and a suggestion from team member B."
"Ramo changed the project's design approach."
The date is gone. The project name is gone. The direction of change is gone. The reasons are gone. Who made the suggestion is gone.
What remains is just the hollow fact that "something changed."
As this repeats, the AI's memory becomes thin, vague, and unusable.
The Session Boundary: Information Isn't Carried Over to Begin With
There's an even more fundamental problem. In many LLM services, memory carryover across sessions (conversation threads) doesn't exist as a mechanism.
The lyrical direction I spent two hours developing in Monday's session. Open a new session on Tuesday, and the AI doesn't even know that conversation happened.
By human standards, this is absurd. Every morning you come to work, your colleagues have forgotten your name. They don't remember yesterday's meeting. They don't know last week's decisions.
Could you advance a ongoing project with someone like that?
This isn't fiction. Right now in 2026, millions of users talk to LLMs daily and hit this wall every day. And many accept it as "inevitable."
But is it really inevitable?
Memory Systems Exist. But—
Of course, the industry hasn't ignored this problem. From 2024 to 2025, various solutions emerged to give LLMs memory.
OpenAI added a "memory" feature to ChatGPT. It automatically extracts facts from conversations like "this person is an engineer" or "they like Python" and saves them as a profile.
Convenient. But users have almost no control over what gets remembered and what gets discarded. If you ask ChatGPT to remember "the project decisions from last week," it won't save them unless ChatGPT judges them worth remembering. The decision criteria are a black box.
The most widely used approach. Past conversations and documents are saved in a vector DB (a database searchable by semantic similarity), and text "similar to" the query is retrieved and injected into the prompt.
A powerful technique. But RAG is "search," not "memory." This distinction is critically important. Precise recall of a specific fact—"the style guide rule 3 we decided on three weeks ago"—can't be fully covered by similarity-based search.
An "AI personalization layer" that rapidly became popular in 2025. It automatically extracts user preferences and facts from conversations and includes conflict detection features.
Convenient. Easy to implement. But while it's good at remembering "fragments of facts," it struggles with understanding relationships between facts. "Decision A is based on premise B"—this kind of logical chain can't be preserved.
The most ambitious project, adapting OS virtual memory management to LLMs. The LLM itself issues Tool Calls to read and write memory. "An AI that manages its own memory."
The philosophy is closest to our approach. But LLM processing capacity is finite. The act of thinking about "what should I remember," "where should I store it," and "when should I retrieve it" consumes the LLM's inference capacity. "Remembering" and "thinking" compete for the same brain.
A Graph+Vector hybrid memory infrastructure for production environments. Excellent at entity extraction and temporal management with enterprise scalability.
Sophisticated infrastructure. But Zep hasn't fully recognized that "remembering" and "organizing" are separate problems. It lacks mechanisms for detecting and resolving logical contradictions. There are also reports of graph "spaghetification" with long-term use.
| Solution | Approach | Strengths | Fundamental Limitations |
|---|---|---|---|
| ChatGPT Memory | Auto-extraction | Convenient, built-in | Opaque, uncontrollable |
| RAG | Vector search | Large-scale ready | Search ≠ memory |
| Mem0 | Fact extraction | Easy to implement | No structure, lacks relationships |
| Letta | Self-management | Flexible | Inference resource contention |
| Zep | Graph+Vector | Scalable | No contradiction resolution |
The Fundamental Oversight Shared by Every Tool
We've reviewed the major memory solutions. Each has strengths, each has limitations.
But there's a fundamental oversight shared by all of them.
Every tool focuses on "remembering." But "remembering" alone doesn't make memory work.
Think of your own brain.
Today, you received hundreds of pieces of information. Things you saw, heard, felt. Most of it, you won't remember tomorrow. Not because your brain is "degrading." Because your brain is "selecting."
Human memory works through "organizing," not "remembering."
- Filter important information
- Resolve conflicting memories ("A is kind" vs. "A got angry at me"—how do we integrate these?)
- Generalize from repeated patterns into rules ("In situations like this, do this")
- Forget information that's no longer needed
This memory's metabolic process—extraction, organization, promotion, generalization, forgetting—is what turns vast information into "usable knowledge."
No existing tool had this metabolic process.
Information gets dumped into a database. Retrieve it with search. That's it. Nobody is "organizing." Nobody is "resolving contradictions." Nobody is "promoting what matters."
Information accumulates. But stays disorganized. So it's unusable.
Imagine an office that just keeps throwing documents into cardboard boxes. They "store" things. But when you need a document, you can't find it. Conflicting memos both stay "valid" in the same box. A year-old policy and last week's new policy sit mixed together.
The core of the LLM memory problem wasn't "forgetting." It was "inability to organize."
Toward the Next Part
When I realized this problem, one question emerged.
"How does the human brain organize memory?"
Cognitive science had a surprisingly precise answer to this question. And it contained hints for software design.
Part 2 will discuss the human brain's memory system—particularly memory consolidation (memory fixation during sleep)—and the idea of recreating it as software.
The multi-layer memory model proposed by Atkinson and Shiffrin in 1968. Working memory as defined by Baddeley in 1974. And the remarkable "memory organization work" the brain does during sleep.
Decades of insights from cognitive science held the blueprint to solving the LLM memory problem.
AI doesn't need a bigger database. It needs a smarter "memory's metabolic engine."
Cognitive science memory models and how to translate them into software. From the Atkinson-Shiffrin multi-layer memory model to MoltMem's architecture.