Yemi ran an eight-person HR consultancy in Manchester. In late 2025, she deployed a client onboarding assistant using a no-code AI platform. The assistant was genuinely capable within a session: it answered policy questions accurately, guided users through benefits structures, and handled complex eligibility questions without needing a human. The problem appeared in the second conversation with the same client. The assistant asked for the company name. It had been given the company name three sessions ago.
Clients stopped initiating new conversations. They preferred sending Yemi a direct email. The assistant had not failed in any conventional sense. Its accuracy within a session remained high. But it had no memory of what had happened before, and without memory, each conversation started as if the client were a stranger.
This is the most common category of AI agent failure in small business deployments: not hallucination, not misrouting, not tool errors. Forgetting.
The digital amnesia problem
The agent memory market reached $6.27 billion in 2025 and is projected to grow to $28.45 billion by 2030 at a 35 percent compound annual growth rate, according to Mordor Intelligence. That growth reflects a specific problem the market is forming around: AI systems that are genuinely capable but functionally limited across time because they cannot remember.
Mem0, a venture-backed company building memory infrastructure for AI agents, raised $24 million in a Series A round in October 2025. Its CEO Taranjeet Singh described the problem in the funding announcement: every agentic application needs memory, just as every application needs a database. The analogy is precise. An application with no database cannot track state, cannot personalise, and cannot improve from past interactions. An AI agent with no memory has the same limitations.
The Mem0 GitHub repository has 58,100 stars, making it one of the most-watched open-source AI infrastructure projects in its category. The volume of developer attention reflects how acutely the problem is felt across teams building AI agents.
Why AI agents forget by design
Large language models are stateless. Each API call receives a context window: a block of text that contains the conversation history, any documents or data passed in, the system prompt, and the current user message. The model processes this block and returns a response. When the conversation ends, nothing is retained. The next conversation starts with an empty context window.
The context window is not memory. It is a working surface. Information placed on that surface exists for the duration of the current call. Information not placed on it does not exist for that call. This is why the HR assistant kept asking for the company name: the company name was given in a previous conversation, which the current call had no access to. It was never on the working surface.
No-code AI platforms reinforce this default. Make.com launched AI agent capabilities in 2025 and these agents maintain conversational context within a session but do not retain memory between separate runs. In n8n, AI agent nodes have in-session memory that is volatile: when a workflow run ends, the memory is cleared. Out of the box, nearly every AI platform ships with this behaviour unless memory infrastructure is explicitly added.
The four types of memory agents can have
The CoALA framework, published by Princeton University researchers in a 2023 paper (arXiv:2309.02427), provides the most rigorous academic taxonomy of AI agent memory. It defines four types.
In-context memory is the working surface: the content of the current context window. It is fast and flexible but limited in size and lost when the conversation ends. Most AI deployments operate entirely on in-context memory.
Episodic memory stores records of past interactions: what happened, when, and in what sequence. A customer support agent with episodic memory knows that this particular user reported a billing issue two weeks ago and that it was resolved on the third contact. That history shapes the current response.
Semantic memory stores general factual knowledge not tied to a specific interaction: company policies, product specifications, regulatory requirements, client preferences. This is knowledge the agent should always have access to, regardless of what happened in recent conversations.
Procedural memory covers learned skills and workflows: how to execute a specific multi-step process, which tools to call in which order, what exceptions to handle. Procedural memory is encoded in the system prompt for simple agents and in fine-tuned model behaviour for more complex ones.
The lost in the middle problem
Adding memory to an AI agent means putting more information into its context window. More information creates a specific risk documented in a 2023 Stanford University paper titled "Lost in the Middle: How Language Models Use Long Contexts" (arXiv:2307.03172).
The paper found that LLM performance is highest when relevant information appears at the beginning or end of the input context. Performance degrades significantly when relevant information sits in the middle of a long context, even in models explicitly designed for long-context use. The effect is robust across model types and task formats.
The practical implication is that memory retrieval matters as much as memory storage. Dumping everything the agent knows about a client into the beginning of every context window is not effective memory management. Effective memory is selective: the system retrieves the most relevant episodic and semantic records for the current conversation and positions them where the model can use them.
Storing memory is not the same as using it. LLMs perform significantly worse on information placed in the middle of long contexts. Memory retrieval and positioning are as important as memory storage.
What forgetting costs businesses
Sinch published a 2026 survey titled "The AI Production Paradox," drawing on 2,527 senior decision-makers across ten countries and six industries. The headline finding was stark: 74 percent of enterprises that deployed AI customer communications agents had already rolled them back or shut them down after launch.
One in three of the deployed agents lacked the customer context needed to deliver an adequate experience. Fifty-five percent of the surveyed enterprises reported that they were custom-engineering cross-channel context preservation because their platform did not provide it natively. Customers who were forced to repeat information rated their experience 76 percent worse than those who were not.
For the Manchester HR consultancy, the cost was measured in adoption rather than customer satisfaction. The assistant failed not because it gave wrong answers but because it gave answers that ignored the accumulated relationship context that clients expected it to carry. The ROI calculation for the assistant depended on clients using it repeatedly. No memory meant no repeat usage. No repeat usage meant no ROI.
The tools that add persistent memory
Several tools have emerged specifically to solve the memory problem for AI agents, ranging from open-source infrastructure to commercial no-code platforms.
Mem0
Mem0 is purpose-built memory infrastructure for AI agents with 58,100 GitHub stars and $24 million raised from Y Combinator, Peak XV Partners, Basis Set Ventures, and angel investors including Dharmesh Shah of HubSpot and Olivier Pomel of Datadog. It provides cross-session memory that stores important information from past agent interactions, manages conflicting or outdated memories, and retrieves relevant context for each new conversation. Mem0 integrates natively into CrewAI, Flowise, and LangFlow, and is the exclusive memory provider for the AWS Agent SDK.
Lindy.ai
Lindy.ai is a no-code AI agent platform designed for non-technical users. It supports persistent cross-session memory by default, storing preferences, prior interaction summaries, and standing instructions in a way that carries across conversations. For a business wanting a client-facing agent that remembers context without developer work, Lindy is the most accessible starting point currently available.
Make.com AI Agents
Make.com AI agents maintain conversational context within a single session but do not retain memory between separate runs by default. Teams that need persistent memory can connect to external databases through the Make integration library, which supports over 3,000 applications, as a workaround. This approach works but requires additional setup and ongoing maintenance.
Zep
Zep is production-grade memory infrastructure for more complex AI deployments. It uses a temporal knowledge graph architecture that stores facts, relationships, and how they evolve over time. It works with any major agent framework and handles the selective retrieval challenge: surfacing the most relevant historical context for each query rather than flooding the context window. Zep is best suited for teams with developer resources and complex memory requirements at scale.
How to choose the right approach
The right memory architecture depends on three questions.
First: does your use case require memory across separate sessions? An agent handling one-off requests with no need for client history does not require persistent memory. An agent handling ongoing client relationships, repeat customers, or cumulative workflows does. The HR onboarding assistant clearly falls in the second category. A one-time document summarisation tool does not.
Second: how complex is the context you need to retain? Simple preferences and account details can be stored in a basic database and retrieved with a lookup. Complex relational history, evolving preferences, and multi-party interaction records need a purpose-built memory layer like Mem0 or Zep that handles conflicts, updates, and relevance ranking.
Third: what is your technical capacity? Non-technical teams should start with Lindy.ai, which ships persistent memory as a platform feature. Teams with light technical capacity should evaluate Mem0, which has integrations for major no-code frameworks and clear documentation. Teams with developer resources who need production-scale memory with audit trails should evaluate Zep.
Where to start this week
The first step is to audit whether your current AI agents actually need persistent memory or whether better use of in-context memory would solve the problem. For many agents, the issue is not that memory is absent but that relevant context is not being passed into the context window at the start of each conversation. If the agent has access to a CRM or a database, pulling the relevant client record at the start of each session may be sufficient without dedicated memory infrastructure.
If your agent genuinely needs cross-session memory and your current platform does not provide it, the path depends on your technical capacity. Non-technical teams: test Lindy.ai on the specific use case where memory failure is most costly. Teams with technical capacity: evaluate Mem0 for a single workflow before extending it to your full agent stack.
The goal is not to give every AI agent perfect recall of everything. It is to give each agent the right context at the right moment. For most SMB use cases, the tools available now are sufficient to solve it without significant infrastructure investment.
Sources
- Mordor Intelligence — Agentic AI Orchestration and Memory Systems Market
- Mem0 — Series A Press Release, PR Newswire, October 2025
- TechCrunch — Mem0 Raises $24M, October 2025
- Mem0 — GitHub Repository
- CoALA — Cognitive Architectures for Language Agents, arXiv:2309.02427
- Liu et al. — Lost in the Middle, arXiv:2307.03172
- Sinch — The AI Production Paradox, PR Newswire, 2026
- The Register — AI Customer Service Bots Rolled Back at 74 Percent of Firms, May 2026
- Make.com — AI Agents
- Lindy.ai
- Zep — Production Agent Memory