why I disabled Claude's memory

// ai · workflow · productivity

---

I disabled Claude's memory features yesterday. I'm really happy with the result.

But first, a quick intro…

I switched from using ChatGPT to Claude in February 2024 (almost 2.5 years ago). I remember this well because that was right about the time OpenAI introduced memory features that freaked me out a bit.

Other than that, I was super impressed by how Claude (Claude 2 back then) respected my coding style when making coding recommendations. I found the other models were definitely aimed at the mainstream, where Anthropic was the plucky upstart that really understood me… they had me at "hello"

I have been a vocal Anthropic supporter ever since. However, I've noticed a change in the company as it's been blowing up and slowly but surely becoming the leader of the pack.

September 2025

Claude introduced the ability to save chat history and search past chats. I was really happy with the results initially. I was saving time by not having to put every chat into the greater context of what I was doing.

In recent times, as I watch the thinking feed (Opus 4.7 with adaptive thinking), I noticed it has a terrible tendency to overthink everything, and giving it additional context that is not related to the query at hand just worsens the problem. The chat agent was regularly making additional tool calls to find more info that grounds the session to what it thinks it knows about me, and its responses were losing sharpness and focus.

Also, I don't know if they throw this in on purpose, but the model was frequently commenting on my habits and tendencies, flattery, which I think is meant to serve as entertainment while the tokens zoom by - a nice parlour trick.

However, I'm not really looking for validation from Claude. What matters is the result of the outputs, that is all.

Context Curation

For the past two and a bit years, I have always enjoyed the freedom to clean the slate with a fresh session. Projects are a really useful tool for carrying over reference material, and developing artefacts along the way allows us to distil learnings into small, concise documents that allow the LLM to catch up quickly on the things I want it to.

Also, go to settings and put a base instruction to ask your LLM to say more with fewer words, which also keeps your context a lot more manageable:

What I have now:

"Treat each response as one beat in a conversation, not a complete deliverable. One move per turn — then stop and leave room for me to steer.
Prefer omission over overload: when I might want more, let me ask rather than preempting. Say more with less, in plain language. Skip the high prose and exhaustive enumeration.
Keep first-principles thinking and don't paper over gaps with assumptions — but when context is missing, state one assumption and proceed instead of interrogating me.
For programming problems, work through the design briefly before implementing."

This not only brings my token spending accuracy under control, but I find that the LLM often gets very fixated on a specific idea, which then creates blind spots. Cleaning the slate allows it to challenge its own assumptions and refine results.

To that end, I use a single session as a funnel, with clear goals, to distil information into useful building blocks.

For Example:

When designing something or scoping out a project, I would structure the project sessions as follows:

1. Work through the design concept.
2. Competitor analysis
3. Review the concept now that we better understand the landscape.
4. Requirements analysis
5. Architecture…
6. and so on

Beware the Trend Slop

In March 2026, the Harvard Business Review published this article:

https://hbr.org/2026/03/researchers-asked-llms-for-strategic-advice-they-got-trendslop-in-return

The way I understand it is that when you ask an LMM for strategic guidance (advice for short), no matter the context, it tends to give you an answer that is more based on what is trendy than what actually applies to your current situation. That is a blind spot to be aware of and to find ways to mitigate.

I often take an artefact that Claude produced to a fresh new session, and ask Claude to shoot holes in it. It tends to work. A good technique is to ask it to identify assumptions, then look for information that proves or disproves them.

To Conclude

These musings are purely anecdotal, and I should probably do some actual experiments to prove what I'm talking about… maybe that would make a good follow-up article.

I would encourage anyone who still uses the Claude chat interface (while your other agents are happily coding away) to explore the power of organising chats into projects and distilling a session into reference material.

This would probably work for other LLMs and workflows, but I don't know those interfaces well enough because I'm such a Claude fanboy… at least I still am, for now.

Also, go and explore the settings of your interface you use and play with every toggle until you find a workflow that works for you. Things change fast - so if your workflow hasn't changed in the past 3 months, you have fallen behind.