After a few months with ChatGPT, people often say: it knows me by now.
That is partly true. Knowing which part matters helps you decide what to rely on and what to check. Features change quickly. This comparison was reviewed on 30 September 2026; availability and interfaces may differ by account and settings.
Three Different Things Are Called Memory#
First is the current conversation. Earlier messages may be in context, but space is limited and a long history need not be processed in full. Closing a window does not automatically delete the stored chat either.
Second are memories and functions drawing on earlier conversations. They can influence a new chat without reliably knowing everything you said before.
Third is your own knowledge base, with defined sources, access rights, maintenance and deletion rules. It can serve several tools. Whether it performs better depends on its implementation.
How ChatGPT Remembers#
OpenAI describes saved memories and reference to earlier chats. Its current help also distinguishes different memory experiences. Availability and settings determine which you use.
ChatGPT can identify information as important itself. You can inspect and edit saved details. Available features may include search, sorting, management and version history. The earlier description of a plain list without those controls was incomplete.
Using memories for later answers does not mean the entire profile and every note are necessarily transmitted in full with each question. The help describes relevant context and explains that a memory summary need not show everything considered from chats. The overview is not complete technical visibility into every request.
Deleting a chat does not automatically remove a separately saved memory from it. Complete removal must account for the relevant storage locations. Temporary chats can help with tasks where you do not want existing memory used.
How Claude Remembers#
Alongside search of earlier chats, Claude offers memory entries that can be organised by topics and edited. Projects can have their own context scope. Check the settings rather than assuming the product name guarantees separation of every piece of information.
Searching the chat archive and using memories are different functions. One can find concrete earlier passages; the other supplies context across conversations. The visible interface does not establish exactly how every entry is loaded internally for each request.
These differences matter, but do not prove one design inherently more honest or reliable. For client work, traceable content, appropriate rights and correction controls matter.
Where Both Need Care#
Memory can summarise incorrectly, associate a detail with the wrong client or keep outdated information. Do not assume that "we now work with B" reliably removes A as the responsible company.
This is a risk, not proof of universal failure in current systems. The earlier success percentage from an unlinked study was not a sound comparison. State important changes explicitly and inspect what was saved.
Migration is possible but limited. Claude documents experimental text-based import. Exports can provide source material. This does not promise a complete one-to-one transfer of chats, relationships, permissions and changes. Review confidential information before import and check the result afterwards.
Sentences or Connections#
Imagine notes saying "Ms Berger works at Sonnenhof" and "Sonnenhof belongs to Nordlicht". Search can find the notes and a model can infer their connection.
A graph stores those relationships explicitly. The person and companies become separate records connected by "works at" and "belongs to". This can help reconcile names and retrieve related information.
It helps only when identities and relationships are captured correctly. A wrongly drawn connection remains wrong. Graph retrieval can miss relevant entries too. It guarantees neither complete answers nor correct reasoning. Text search, semantic retrieval and graphs can complement each other.
Dates Help When Their Meaning Is Clear#
When Ms Berger changes companies, a system can end one relationship and create another. This can reconstruct a past situation if the change was captured.
A storage date is not automatically a validity date. Something entered today may describe last March. Source, valid period and capture time may all matter. A newer statement does not automatically resolve a contradiction correctly.
Retrieve the Relevant Page#
A large collection should not be attached indiscriminately to every request. Search can select relevant information, including without a graph. Current services use different selection methods. The claim that ChatGPT and Claude always include everything was too broad.
Claude Code offers a concrete example. Its documentation describes automatic notes and a MEMORY.md index whose first 200 lines or 25 KB are loaded. Other topic files are read as needed. Larger contents remain stored even when not fully loaded into this initial context.
That is a product-specific rule, not a universal law. Selection can overlook important material. Files such as CLAUDE.md or AGENTS.md need clear scopes and consistent instructions. An external knowledge base needs maintenance too.
What We Aim to Achieve with Our Own Memory#
Our approach combines search with explicit people, companies, projects and relationships. The goal is traceable retrieval and reviewable changes. Test whether that meets your needs with real questions.
Uncertain contradictions should be flagged. Two similar proposals should remain separate events. These are important requirements, but automation cannot guarantee detection of every error.
Infrequent retrieval does not automatically make information obsolete. Conversely, a client potentially returning does not justify unlimited retention of personal data. Purpose, legal basis, retention and deletion must be defined. Our earlier statement that people and client projects never expire was too absolute.
Imports also need review. They may extract selected information from exports, but format, content and supported features determine the result. Neither complete transfer nor correct matching should be assumed.
What You Can Inspect#
A map of records and relationships can expose missing or wrong connections. A time view can show earlier states if the system really retains and dates them correctly. Neither replaces source checks.
The claim that no major service offers an earlier state was incorrect: ChatGPT's help describes version history for available memory features. That is not necessarily a complete historical company map. Compare actual functions rather than metaphors.
Where Extra Memory Adds Little#
Occasional rewriting may not require a separate knowledge base. Extra context takes effort and can distract. For a fresh opposing view, a chat without established preferences may help, though its answer still needs checking.
There is no universal rule for company notes. "Use informal address" may help directly; a company description can be essential for one task and irrelevant for another. Test it on your work rather than generalising from one experiment.
What Matters After a Year#
Age alone makes memory neither useful nor unusable. Current sources, visible changes, appropriate retrieval and regular maintenance matter. Measure recurring questions before treating a large collection as success.
Start with your service's settings. Correct inaccurate details and check sources for important questions. Repeatedly explaining context, or needing controlled shared knowledge, is a good reason to test a knowledge base of your own.
Sources and review date#
Reviewed on 30 September 2026.
