Every time you start a fresh chat and type “just so you know, I’m a freelance copywriter working mainly with B2B SaaS clients…” — that’s a tax. A small one, but it compounds. If you’re using AI seriously, you’re paying it dozens of times a week.
The tools that claim to fix this are all slightly different under the hood, and the difference matters more than the marketing suggests. Here’s what’s actually going on with each approach.
What “memory” actually means in each tool
ChatGPT’s memory feature — visible under Settings > Personalization — stores discrete facts the model picks up from conversations and surfaces them in future sessions. It’s not a transcript. It’s more like a sparse notes file. The model decides what’s worth saving, which means it keeps “user prefers bullet points” but tends to drop nuance like “user is trying to move away from client work and into productized services.” You can read and edit what it’s stored, which is genuinely useful — but most people never look at the list and then wonder why the model keeps giving generic advice months later.
Claude has no persistent memory between sessions by default. Each conversation starts cold. What it does have is a very long context window — up to 200k tokens on Claude 3.5 and newer models — which means if you paste a 2,000-word background document at the top of a session, it holds onto it with unusual fidelity for the entire conversation. That’s a different trade-off: high-quality within-session recall, zero cross-session persistence unless you wire up something external.
Gemini sits somewhere between the two, depending on which Google integrations you’ve enabled. It can surface your calendar entries, emails, and Drive files — which can feel like memory but is really just retrieval. It’s pulling existing documents, not learning your preferences. Useful for specific tasks, irrelevant for others, and worth keeping distinct in your head from actual preference memory.
The problem with model-managed memory
Handing the model control over what to remember introduces a specific failure mode: it optimizes for what looks memorable, not what’s actually load-bearing for your work.
Say you tell ChatGPT mid-conversation that you’re working on a project for a legal-sector client who’s unusually risk-averse about AI-generated content. The model might store “user works with legal sector clients.” It probably won’t store “this client wants outputs that read like they weren’t AI-written and will ask about it.” The second thing is the part that changes the actual output — and it’s gone the moment the session ends.
This is compression, which is what any summarizer does. The practical consequence is that built-in memory works reasonably well for stable, high-level preferences (tone, format, role) and poorly for project-specific nuance that actually matters for a given task.
What actually works: the boring answer
A plain text file.
Keep a “context block” — a few hundred words covering who you are, what you’re working on right now, and any standing preferences that should affect output. Paste it at the top of any session where it’s relevant. Update it when your situation changes.
This sounds worse than automated memory. In practice it’s better, for three concrete reasons. First, you control exactly what’s in it — no silent compression. Second, it works identically across ChatGPT, Claude, Gemini, or anything else, so there’s no lock-in. Third, writing and maintaining it forces you to actually articulate your context, which tends to improve the quality of what you get back regardless of any memory features.
The version worth keeping has three sections: a two- or three-sentence description of who you are and what you do; a short list of current projects or focus areas; and a list of output preferences covering format, tone, and what to avoid. Keep the whole thing under 300 words so it doesn’t eat a meaningful chunk of your context budget.
Where dedicated memory tools fit in
Third-party tools — Mem, Notion AI, some custom GPT setups — take a more structured approach to storing and retrieving past conversations. They tend to outperform built-in model memory for one specific case: ongoing work over weeks or months where you want to query across sessions, not just carry preferences forward.
If you’re running a long research project or using AI as a running sounding board for a business problem that’s evolving, something that indexes your conversation history and lets you retrieve it is genuinely useful. Where it breaks down is anything requiring real-time context about what you’re doing right now — these tools surface historical information, they don’t understand the current session.
One thing worth being honest about
AI memory features are still unreliable enough that you shouldn’t treat them as a system. ChatGPT’s memory has improved, but it still drops things it shouldn’t and occasionally surfaces stale context as if it were current — a project you finished months ago, a preference you’ve since reversed. Confidently wrong is a worse failure mode than blank.
Use the features, but treat them as a lightweight assist. If the context actually matters for the output — and for anything important it does — check that the model actually has it right before you let it run unsupervised.
The actual decision
If your AI use is mostly one-off tasks with no real continuity between sessions: don’t bother with any memory setup. Paste context when it matters.
If you use one tool consistently and want it to remember stable preferences without re-pasting every time: ChatGPT’s built-in memory is fine for that. Go look at what it’s actually stored — it’s probably not what you’d expect.
If you want cross-tool consistency or you’re doing serious ongoing work across sessions: the plain text context block will serve you better than any automated system right now. It takes five minutes to set up.
The question worth sitting with: how much of your current AI friction is actually a memory problem versus a context problem? They look similar but have different fixes — one is about persistence across sessions, the other is about the quality of what you’re feeding in right now. Most people reaching for memory solutions would get more mileage from improving what they’re putting in the prompt.
