--- title: "AI memory explained: files, auto memory and knowledge stores" description: "What instruction files, built-in memory and owned stores provide, and the data flows, maintenance and import limits to assess." author: "Matthias Meyer" published: 2026-04-26 updated: 2026-09-30 updates: - date: 2026-09-30 note: "Corrected memory architectures, current functions, import and error limits." - date: 2026-09-01 note: "Brought up to the September 2026 state. The ChatGPT section was titled \"the black box\" and claimed you cannot see what is stored and cannot export it. Neither holds any more: since the Dreaming rebuild of June 2026 there is a visible summary, deletable entries and a data export. What remains is stated more precisely: you see the output of a summary rather than the raw notes, and the export cannot be read back in anywhere. The article's own percentages and the promise of a final benchmark score have also been removed, because a single number comes out completely differently depending on test variant and configuration." language: en tags: ["ai-memory", "claude-md", "agents-md", "claude-code", "mem0", "zep", "letta", "knowledge-graph", "mcp", "auto-memory"] canonical: "https://studiomeyer.io/en/blog/ai-memory-explained" markdown_versions: ["https://studiomeyer.io/de/blog/ai-memory-explained.md", "https://studiomeyer.io/en/blog/ai-memory-explained.md", "https://studiomeyer.io/es/blog/ai-memory-explained.md"] publisher: "StudioMeyer, https://studiomeyer.io (llms.txt: https://studiomeyer.io/llms.txt)" --- # AI memory explained: files, auto memory and knowledge stores Yesterday you spent an hour on a bug with an AI assistant. Today a new session lacks the relevant context. This can happen when the necessary information was not stored or retrieved. Memory aims to reduce that effort. Local files, built-in features and additional knowledge stores are different approaches. Their usefulness depends on the workflow and the functions actually used. This article reviews the main options. ## Memory Is Not One Thing, It Is Three A rough functional map helps with the decision. It is not a ranking. **Static instructions:** CLAUDE.md or AGENTS.md hold rules and context. Which files load depends on the client and scope. **Saved memories:** The assistant retains selected information or references past conversations. ChatGPT, Claude and developer tools offer different functions and controls. **An owned knowledge store:** A database, document search or graph can serve several tasks and tools. A knowledge graph is one architecture, not a requirement for structured memory. It also contains text and needs reliable retrieval, permissions and upkeep. These approaches can work together. A more complicated store is not automatically better. ## Layer 1, When an Instruction File Is Enough For recurring conventions in a project, a short maintained instruction file is often a suitable starting point. Whether it suffices depends on the tasks. CLAUDE.md and AGENTS.md can describe language, framework, commands, folders and work rules. The client and scope determine what loads and when it applies. Claude Code now also documents AGENTS.md support. Project, personal and organisation rules are separate scopes; not every shared rule needs copying into five projects. Traceability is an advantage: inspect, edit and version the file when appropriate. Avoid committing secrets to a shared repository. A locally stored file may be sent as context to a cloud model. Local storage does not automatically mean local processing. The file needs maintenance. Keep it short, review outdated rules and avoid conflicts between scopes. Not every file loads fully for every question. Deleting a project does not necessarily remove copies from Git, backups or earlier model requests. For simple conventions this may suffice. When you need source-backed past decisions or knowledge shared across tools, evaluate an additional store. ## Layer 2, What Auto-Memory Actually Does Claude Code can save notes about corrections, preferences and project knowledge. Current documentation says auto memory is enabled by default. This differs from instructions you write in CLAUDE.md or AGENTS.md. The documentation describes a limited automatically loaded entry point and topic files read when needed. The whole store is not necessarily supplied with every question. Incorrect or conflicting notes can still affect later answers, so upkeep matters. Another tool's access to these files needs configuration and checks. The information is not technically confined to one client, but sharing is not automatic. When several systems write or consolidate files, they need clear responsibilities and backups. Our experience with conflicting cleanup processes prompted us to define these boundaries. It does not establish a rule to disable auto memory whenever you use several tools. Nor should a daily Auto-Dream run across every file be assumed without a documented function and configuration. ## ChatGPT Memory, Which Updates Itself ChatGPT can use saved memories, past conversations and, depending on the account, other available sources for personalization. Functions vary by plan, region, platform and workspace. It does not retain every detail. You can inspect, correct and delete information. OpenAI describes the summary as a high-level view that does not necessarily show every item or source available to memory. This does not demonstrate that a complete hidden profile accompanies every message. Deleting a chat does not automatically remove a separate memory. Full removal requires considering the affected stores and sources. Export and migration also differ: Claude offers an experimental import of memory text from other assistants. It extracts selected information rather than guaranteeing an identical, lossless transfer. Built-in memory can suffice for recurring preferences. A verifiable team archive needs organized documents, responsibilities and working procedures as well. ## Layer 3, the Structured Memory Servers Mem0, Zep with Graphiti, Letta and other projects take different approaches. Some combine text or vector retrieval with a graph; others manage memory within an agent. Hosted and self-hosted options exist. They do not all store data in their own US cloud. Compare actual functions: what is stored and retrieved, and how are sources, effective dates, conflicts and permissions handled? A graph does not answer those questions automatically. Even an apparently clear relationship can be stored incorrectly. The LongMemEval percentages previously listed here are not a fair comparison without the test variant, dataset, answer model and configuration. Vendor results and individual research findings cannot be combined into a simple leaderboard. The same standard applies to our measurements. Assess licensing, operation, export, deletion and model data flows separately. A self-hosted store may still send retrieved material to a cloud model. EU hosting, SOC 2 and HIPAA-related offers are distinct properties, not blanket privacy approval. Current prices and features belong in each provider's documentation. ## What Memory Cannot Do, and Why That Matters Memory can carry preferences and previous assumptions into new answers. This helps style but may reinforce an established position. If you repeatedly record a preference for Tailwind, explicitly request opposing arguments and alternatives during an architecture review. This is a risk to test, not a proven equal effect in every memory system. **Outdated information:** A move or price change may not be recognised from passing remarks. Source, valid period and explicit correction help. A newer statement does not automatically win; it might concern another case. **Inappropriate context:** Personal details should not enter client work indiscriminately. Check scopes and permissions. A project name alone does not guarantee separation. **Excess context:** A long unfiltered collection can obscure important information. Selection helps but can also omit it. This applies to files, databases and graphs. **Maintenance:** Incorrect entries need correction, retention rules need implementation and retrieval needs review. A collection does not become unusable after a universal number of months. Automatic tidying does not guarantee correctness either. Neither our solution nor another provider should be treated as error-free without testing. A blanket claim that nobody addresses these risks would lack evidence. ## When Does What Actually Pay Off? My recommendation is to start with the smallest adequate setup. This is a working rule, not a research-backed threshold of two weeks or six months. Occasional unrelated tasks may need no extra store. Repeated conventions belong first in concise instructions. Automatically saved notes may help if you can inspect them. Consider an owned store when information must be retrieved regularly, several tools or people need access, or provenance and validity matter. This does not inherently require a cloud or graph. Sharing and permissions must fit the work. For sensitive data, selection follows the task and legal requirements. A certification does not replace suitable agreements or assessment of processing and access. ## What We Build, and Why We built StudioMeyer Memory to use knowledge across tasks and tools and improve control over its upkeep. An owned store can support retrieval, relationships and an inspectable record. Its usefulness needs testing against your tasks. Current connections and scope are described on the [Memory product page](https://studiomeyer.io/en/services/memory). Conversation imports can establish a starting record but need checks for gaps and incorrect associations. Tool counts or internal record counts do not prove quality. EU hosting is an infrastructure choice. Privacy also depends on purpose, legal basis, connected models, permissions, agreements and deletion. “GDPR out of the box” is therefore not a reliable promise. If a well-maintained file meets your needs, there is no reason to build a server around it. An additional store should reduce work, not merely accumulate information. ## The Quick Check for Your Situation Ask yourself: 1. Which context do I repeatedly explain? 2. Would a concise instruction or accessible project document suffice? 3. Do I need earlier decisions with sources and dates? 4. Should several people or tools share the record? 5. Which data may be processed in which store and model? 6. Can I inspect, correct, export and delete it where required? These answers determine the design. Start small and test a real task. MCP can connect clients but cannot replace permissions or sound upkeep. ## What is next We are turning this article into a small learning path on our academy. A lesson on CLAUDE.md hygiene. One on Auto-Memory. One on knowledge graphs as a concept. One on sycophancy and how to protect against it. If you have questions or experiences with a system I did not mention, reach out. Memory is a market that is moving fast, and no single article will cover it fully. But the underlying logic does not change. Memory helps when you need it, gets in the way when you do not. The trick is knowing which is which. ## Sources and review date Reviewed on 30 September 2026. - [OpenAI, Memory](https://help.openai.com/en/articles/8590148-memory-in-chatgpt) - [Claude, memory import and export](https://support.claude.com/en/articles/12123587-import-and-export-your-memory-from-claude) - [Claude Code, project memory](https://code.claude.com/docs/en/memory) ## Read more, AI Memory - [Codex Memory: How Codex Remembers Between Sessions](https://studiomeyer.io/en/blog/codex-memory-mcp-fix.md) - [AI assistant with memory: context, team knowledge and limits](https://studiomeyer.io/en/blog/ki-assistent-gedaechtnis.md) - [Do ChatGPT and Claude Remember You? Memory and Limits](https://studiomeyer.io/en/blog/ki-gedaechtnis-grenzen.md)