Why memory is the most important unsolved problem in AI

The AI industry has made extraordinary progress on reasoning, generation, and tool use over the past three years. Models can write code, analyse documents, hold nuanced conversations, and coordinate complex multi-step tasks. The capabilities that seemed like research milestones in 2023 are production features in 2026.

Memory is the exception. Most AI systems in production today have no persistent memory. Every conversation starts from zero. Every session is a first meeting. The agent that helped you yesterday has no recollection of it today. This is not a minor gap in an otherwise complete technology. It is a structural limitation that constrains what AI can become.

What memory means in this context

Memory for AI agents is not the same as storage. It is the ability to persist, organise, retrieve, and reason over information across sessions and over time. That includes extracting what matters from interactions, storing it in a structured and retrievable form, surfacing it when relevant without surfacing it when not, tracking how facts change, resolving contradictions, and linking fragmented references to the same entity.

This is the agent memory problem. It is distinct from RAG, which retrieves documents from a static corpus. It is distinct from context window management, which handles information within a single session. Memory is the layer that makes an agent genuinely useful over time rather than merely capable in the moment.

Why it is harder than it looks

Memory sounds simple in the abstract. Store what the user said, retrieve it later. In practice, every component of that description contains a hard problem.

Retrieval at scale

A memory system with ten facts in it can return the right one on every query through brute force. A memory system with thousands of facts across months of interactions needs to find the relevant ones in a large, noisy corpus where many facts are semantically adjacent to any given query. Semantic collapse means that as the store grows, similarity-based retrieval becomes an increasingly unreliable guide to relevance. The retrieval problem gets harder, not easier, with scale.

Temporal reasoning

Facts change. A user's role, preferences, decisions, and circumstances all evolve over time. A memory system that treats all stored facts as equally current will surface stale information alongside current information with no signal about which is which. Memory drift is invisible and compounding: the system does not know it is returning outdated information, and the user has no way to tell from the output.

Entity resolution

Users refer to the same things differently across conversations. "John," "my manager," and "the person who reviewed the proposal" might all be the same entity. A memory system that treats each reference independently accumulates fragments it can never connect, producing a permanently incomplete picture of any entity mentioned across multiple sessions.

Relevance judgement

Perhaps the hardest problem of all: knowing when a memory is worth surfacing and when it is not. Most memory implementations surface everything above a similarity threshold, which produces the forced, performative recall that has given users a low opinion of AI memory as a capability. Good memory is invisible when it is working. The user does not notice the memory, they notice the answer is better. Building a system that makes this judgement correctly requires multi-signal scoring, re-ranking, and coherence checking rather than a single similarity threshold.

Multi-session synthesis

Some questions can only be answered by assembling information from multiple past sessions that were never stored together. A decision made in February, a constraint mentioned in April, and a preference stated in June might all bear on a question asked in September. Finding and assembling these fragments is a retrieval architecture problem that vector search alone cannot solve.

Why the large labs have not solved it

The large labs, OpenAI, Anthropic, Google, Meta, are focused primarily on model capabilities: reasoning, generation, multimodal understanding, tool use. Memory is an infrastructure problem that sits outside the model. It requires retrieval architecture, storage design, and production engineering rather than training breakthroughs.

The labs have shipped some memory features in their consumer products, and the results are mixed. Users frequently report that memory feels forced, irrelevant, or unreliable. This is consistent with the difficulty of the problem: bolting memory onto a model as a feature is a different exercise from building a memory architecture from the ground up.

The result is that memory leadership currently sits with infrastructure companies rather than model providers. M-1, Exabase's memory engine, leads both major memory benchmarks at every evaluated scale, using a cheaper model than every competitor. The gap between M-1 and the next best system widens at the largest and hardest scale (10M tokens on BEAM), suggesting that the architectural advantage becomes more pronounced as the problem gets harder.

Why it matters more than people think

Memory is not a feature. It is the capability that unlocks a different class of AI application.

Without memory, personalisation is impossible. An agent that does not know its user cannot adapt to them. Every interaction is generic.

Without memory, improvement over time is impossible. An agent without memory offers the same quality of assistance on day one thousand as on day one.

Without memory, trust cannot be built. People invest in tools that learn. They develop habits around systems that remember. They build workflows that depend on continuity. An agent that resets to zero on every session cannot become the kind of tool people build their work around.

The gap between what AI can do in a single session and what it can do across time is almost entirely explained by memory. Closing that gap is the most consequential infrastructure problem in AI right now.

Where it stands today

Memory is not fully unsolved. Significant progress has been made, particularly in the past year. M-1's results on LongMemEval (96.4%) and BEAM (68.0% at 10M tokens) demonstrate that reliable memory retrieval is achievable at production scale, with the right architecture.

But BEAM's category-level results also reveal what remains open. Multi-session reasoning at 10M tokens, where the answer requires assembling fragments from across a massive corpus, is still the hardest problem in the benchmark. M-1 scores 9.6% on this category at 10M scale. Every other system scores lower or has not published results. This category represents a challenge for the field as a whole, not a failure of any individual system.

The problem is solvable. Parts of it have been solved. The rest is a matter of architectural progress rather than fundamental limitations. The companies and teams working on it now are building the infrastructure layer that will determine whether the next generation of AI agents is genuinely useful over time or merely impressive in the moment.

Exabase is built around this problem. The Memory API is the result of building memory infrastructure for Fabric's production users before offering it as a platform for other developers. The research documents where M-1 stands today and where the open challenges remain. The docs cover how to get started.

FAQs

Is memory really "unsolved" if M-1 scores 96.4% on LongMemEval?

LongMemEval tests memory at roughly 115K tokens. At that scale, M-1's results are near-ceiling. At 10M tokens on BEAM, the hardest categories, particularly multi-session reasoning, remain genuinely open challenges for every system in the field. Memory is partially solved and the unsolved portion is the most consequential for production applications at scale.

Why have the large labs not prioritised this?

The labs are focused on model capabilities: reasoning, generation, tool use. Memory is an infrastructure problem that sits outside the model. It requires retrieval architecture and production engineering rather than training breakthroughs. The labs may eventually build memory infrastructure, but currently the leading results come from specialised infrastructure companies.

What is the difference between AI memory and RAG?

RAG retrieves documents from a static corpus at query time. Memory persists facts, preferences, and state across sessions and evolves as information changes. RAG answers "what is in this knowledge base." Memory answers "what do I know about this user and how has it changed." They are complementary, not interchangeable.

What makes memory harder than retrieval?

Retrieval finds similar things. Memory requires knowing what is currently true (temporal reasoning), linking fragmented references (entity resolution), resolving contradictions (knowledge update), determining what is worth surfacing (relevance judgement), and assembling information from multiple past sessions (multi-session synthesis). Each of these is a distinct engineering problem on top of retrieval.

What is the BEAM benchmark?

BEAM tests memory at 100K, 1M, and 10M token scales, the equivalent of months to a year of daily conversations. At 10M tokens, no context window can hold the corpus. The only path to a good score is retrieval architecture that genuinely works. See the BEAM research paper.

What is multi-session reasoning and why is it so hard?

It requires answering a question by assembling fragments from multiple separate conversations that were never stored together. At 10M tokens, the relevant fragments are scattered across a massive corpus of irrelevant conversation. This is the hardest retrieval problem in any current benchmark and remains an open challenge.

Does memory require a specific model?

No. Exabase is model-agnostic. M-1 achieved its benchmark results with Gemini 3 Flash, but the memory API works with any model. The retrieval architecture is what determines memory quality, not the model choice.

How does Exabase approach this problem?

M-1 treats memory as reconstructive, drawing on cognitive science frameworks for episodic and semantic memory. The retrieval pipeline combines semantic similarity, lexical precision, temporal salience, importance scoring, and cross-memory coherence. The architecture is documented in the research papers.

Will bigger context windows solve the memory problem?

No. Research consistently shows that model performance degrades with context length. Larger windows provide more capacity for irrelevant tokens that the model cannot reliably attend to. The case for precise retrieval gets stronger as windows get larger. See context window overflow.

Where can I learn more?

The what is agent memory hub page links to every concept, problem, and comparison page in the content cluster. The research page covers the full benchmark methodology and results.