Every Claude Code Memory System Compared (So You Don't Have To)

Mem0 is a cross-tool memory layer well-funded and widely used by developers (100,000 from large companies), designed to be more production-ready than Open Brain, and can be set up in less than a minute

“there is an alternative that I quickly wanted to mention here called Mem0, and it's another cross-tool memory layer. It's well-funded, it's really well-marketed, and widely used already by developers. And you can see, like, used by 100,000 developers from all of these large companies... the setup is designed to actually just get you started in less than a minute”

When you search for the best Claude code memory system, you get buried by many competing tools (Mem Zero, Carpathia's Obsidian Wiki, Open Claude Hermes, Mem Palace, Light Rag, Claude Mem, Claude Mem) but these aren't actually competing tools—they're different ways of handling memory depending on your specific use case.

“Right now, if you search for the best Claude code memory system, you get absolutely buried. Mem zero, Carpathia's Obsidian Wiki, Open Claude Hermes, Mem Palace, Light Rag, Claude Mem, Claude Mem. It's getting ridiculous to keep up with it. But after spending hours researching it, these aren't competing tools. They're different ways of handling memory depending on your use case.”

Andre Karpathy's LLM Wiki pattern uses two folders (raw for source documents, wiki for Claude-owned output) with plain markdown files, where Claude reads from raw but never writes to it, and maintains the wiki structure and cross-references automatically without user intervention.

“What makes it a bit different is that it's a pattern for using Claude to build a living wiki about any topic that you care about. So, you literally just create two folders inside your project. You create a raw folder where you drop your source documents, which are articles, your reports, your podcast transcripts, YouTube video transcripts, PDFs, and Claude is going to read from this folder, but never actually write to this folder. And then you have a wiki folder, which Claude owns completely. So, it's going to write every file, maintain the structure, update cross-references as you go, and you never actually write yourself to the wiki.”

Level 3 uses semantic vector search (via MemSearch plugin from Zilliz) to chunk documents into semantic vectors, allowing search to understand meaning behind questions and stored information rather than relying on keyword matching, with automatic injection of top 3 semantic matches into every prompt via a user-prompt-submit hook.

“it basically chunks your documents, everything you write, into semantic vectors. So, that when we're actually searching for information, it's understanding the meaning behind our questions and the meaning behind the stuff we've stored, rather than just using keyword search. But it takes that one level further and it applies a concept we've already seen already, which is actually hooks. So, it auto-injects the top matches into every single prompt. So, as soon as we write a prompt, it's going to use a hook called user prompt submit, and it's going to feed in our top three semantic matches from our memory files directly into the context for Claude Code.”

Mem Palace stores information using two separate databases: an SQL database that tracks entities and relationships, and a Chroma vector database that stores every conversation as searchable chunks in drawers, with silent background indexing via Claude Code hooks on session end, pre-compaction, and write events.

“it effectively uses two separate databases, an SQL database that tracks entities and relationships between them, and a Chroma DB or a Chroma vector database that stores every single conversation as searchable chunks as we've seen in the drawers. And similar to some of the previous levels, it's also using background hooks to silently store and index information. So, filing and indexing happens silently through claw code hooks on session end. So, we've seen the session start hook. On session end, on pre-compaction. So, if you're about to compact, it will also fire the hook. You write it and the palace is going to basically put that information somewhere behind the curtain.”

Recall is a hosted service that automates Carpathy's LLM Wiki pattern, installing as a browser extension to save articles and podcasts without requiring Obsidian setup, with built-in AI chat and MCP access

“Recall is a hosted service that's basically Karpathy's wiki, the LLM Wiki done for you. So, you install a browser extension, you can save articles, you can save all your podcasts, PDFs, etc. And Recall summarizes, tags, and auto builds that knowledge graph for you without the the setup headache... You don't need to configure or download Obsidian, and it's basically just available through a login.”

Anthropic is working on an unreleased always-on daemon called Kairos (found in accidentally leaked Claude Code source code) that watches your project continuously, decides what's worth remembering, consolidates old notes, and is designed to solve the context rot problem and improve the native memory system.

“when the Claude code source accidentally leaked a couple of weeks back, people found references to a system or framework called Kairos and had a look through the code and it's an unreleased always-on demon that watches your project continuously, decides what's worth remembering, and then consolidates old notes while you sleep. So, it's not in any public build yet, but for sure, Anthropic in the background is working on features to actually solve this context rot problem and improve the native memory system.”

Every memory system for AI assistants is answering exactly the same core question: when you give Claude a task, how does it pull the right context at the right time.

“every memory system is answering exactly the same question. When you give Claude code a task, how does it pull the right context at the right time?”

Level 6 (Open Brain) is an infrastructure layer for thinking that runs one database, one chat channel, and allows any AI tool (ChatGPT on phone, Claude Code at desk, Cursor, Claude Desktop) to plug into it and see the same memory automatically in real time, making it the most future-proof and portable system.

“Level six is a method I've seen called Open Brain by Nate Jones...And the intent of this is actually allow you to switch between AI tools constantly, like ChatGPT on your phone, code code at your desk. And it's the only system today where all of your different AI that you're using are able to see the same memory automatically in real time. So arguably, it's the most future-proof and most portable because actually your memory for this is going to live in a Postgres database that you own. So when the next big AI tool drops in 6 months, say, you just connect it to your existing Postgres brain, and there's no actual migration needed...It's an infrastructure layer for your thinking. So it's one database, one chat channel, and any AI you use can plug into it, which makes it super powerful if you work across apps.”

MemSearch is built by Zilliz, the team behind one of the most popular open-source vector databases, and ports Open Claude's memory architecture into a Claude Code plugin with a two-line install while maintaining a markdown-first philosophy where everything is readable and portable.

“And it brings us to a framework called MemSearch, which is actually a plugin. And it basically ports that framework from OpenClaude over to Claude Code. So, it's built by a company called Zilliz. And that's a team behind one of the most popular open-source vector databases out there...it's got all the benefits of the OpenClaude method, which is the markdown-first philosophy. So, everything we can go, we can see, we can read through, and we can port it to different systems...it works inside Claude Code, too.”

Claude.md is a plain markdown file inside a project folder where you store rules, brand information, client details, and coding style, and it is always loaded into every single Claude Code session in the terminal, functioning like a system prompt given to Claude.

“it's effectively a plain markdown file inside your project folder where you store your rules, stuff about your brand, things about how you work with clients, coding style, and it's always loaded into every single session you pull up inside the terminal. You can think of it like a note that is just at the start of the context, almost like a system prompt that you're giving to Claude.”

Recall has three primary disadvantages compared to Carpathy's LLM Wiki: (1) data lives on Recall's servers, not owned by user, (2) designed for content consumption, not operational memory (e.g., can find 'which video discussed X' but not 'what did we decide about client Y in March'), and (3) carries pricing implications

“Why I think it's not as good as Carpet these pattern for most people. The first one is ownership. So, you don't actually own the data. Recall lets you export everything to markdown if you want to, but while it's live, it's actually live on their servers... And secondly, Recall seems to be built for content consumption, not operational memory. So, it would be really good for a use case like I watched 40 YouTube videos on core code. Which one talked about memory, and what did they mention about memory? But, it's not necessarily designed for what did we decide about client X's landing page and the conversion rates back in March?... And then thirdly, because it is a off-the-shelf software, there is a pricing implication too with that.”

Mem0's main downside compared to Open Brain is that your data lives on their servers permanently, whereas with Open Brain you own the Superbase project and can export the whole Postgres database whenever you want and control access to multiple different tools.

“But again, one of the downsides here is that your data is going to live on their servers permanently, whereas something like OpenBrain, you actually own the Superbase project, and you can export the whole Postgres database whenever you want and control access to multiple different tools.”

Most users experience repeated information to Claude because the current memory system is not reliably feeding memory files in at the right time, which motivates Level 2 improvements.

“You might have noticed this if you keep repeating information to Claude that you think it should already know. And that is where level two comes in. So, level two improves the reliability of feeding those memory files in at the right time, as well as improves the inbuilt structure inside Claude.”

Open Brain's main downside is significant complexity and setup time compared to other levels, but it has AI-assisted setup guides (including video walkthroughs by Nate Jones and companion prompts to migrate from existing Claude Code systems) and costs approximately 10-30p per month on Superbase's free tier.

“You have a full setup guide, which I'll link down below, or a video walkthrough on YouTube from Nate himself, as well as some companion prompts that help you migrate your memories from existing Claude code memory systems over to this Open Brain...and it is probably overkill for most people, but I'll link to Nate's guide, which is written for non-technical people, and you'll get through it if you want to do the setup in about 30 to 45 minutes...one of the downsides compared to the other levels we've seen, one we're still storing away from our laptop, so it's not locally held data. It is secure in a database, but it effectively means we have to query an external-facing database every single time, which is going to add some latency to our process...although OpenBrain, as we mentioned, is going to be like 10 to 30 p per month on the free tier of Superbase.”

Level 2 of memory systems improves reliability by using a structured memory approach rooted at .claude/memory with subdirectories for general.md (cross-project facts and preferences), domain-specific knowledge (topic markdown files), and tools (individual tool config files like slack.md), injected automatically via a Claude Code session start hook.

“Structured memory system rooted at .claude/memory. We already saw our automatic memory system inside Claude, but instead, we've got a structure here. We've got memory.md, general.md, which stores all cross-project facts, preferences, environment setups, but then we've also got domain-specific knowledge, where we effectively have domain/topic, and each topic can have its own markdown file with domain specific knowledge, one file per topic, and the same with tools. We can have individual tools like slack.md with a specific tool config”

Even at Level 4, everything is still locked to your local machine, so information from conversations in ChatGPT, research done on your phone, or brainstorming on other tools is not accessible, which is why Levels 5 and 6 exist to build knowledge bases across multiple tools.

“But, even now, everything's still locked into your machine. So, what about the research you did in ChatGPT last year? The brainstorm you did on your phone the other day? That's where level five and six come in, and they're totally different. They're not about remembering conversations anymore. It's about building up a knowledge base across tools.”

Karpathy's LLM Wiki pattern produces a knowledge graph visualization in Obsidian that shows how different resources interconnect, and it's best suited for use cases like having a pile of bookmarked Slack threads, YouTube videos, client notes that you never go back to, or creating a Wikipedia on a specific subject topic.

“And effectively, inside of Obsidian, which is a separate software that you can download for free, once you compile all of your resources, you'll get this knowledge graph of how your different resources interconnect...It's best in use cases where you have folder called save for later, which is basically a pile of bookmarked Slack threads, YouTube videos, client notes that you never go back to, that you want to see the interlinking relationships with. And as it sounds, effectively create a Wikipedia on a specific subject topic.”

The memory.md file should function as an index pointing to separate files rather than storing all memories directly; for example, it contains 'feedback on project structure' which points to feedback_project_structure.md, keeping the index under 200 lines and only loading specific files when needed.

“It's following exactly what we talked about, which is don't just dump more than 200 lines in a single file. Always always reference a separate file and treat the memory.md or the Claude MD as an index and then point to another file.”

Level 5 is for building interconnected knowledge bases across consumed information (articles, videos, podcasts, client notes) rather than operational memory, and should only be added if you regularly consume information and want to keep and connect it over time rather than consuming content casually.

“Level five is totally different, actually. It's not anymore about remembering conversations. The Mem Palace is kind of the best, the pinnacle of remembering word-for-word conversations. So, now it's about actually, if you want to build a knowledge knowledge base of interconnected information. So, you want to add level five if you regularly consume information that you want to keep and connect over time. So, think about articles, videos, podcasts, client notes, anywhere where you're actually just reading a lot of information and losing track of what you actually learned. You can connect that all into a interconnected knowledge base...I'd skip this if you consume content casually and don't need to come back to it later.”

Context rot is the inability for AI and LLM models to recall 100% of information that has been loaded in as the amount of context loaded increases, and as a general rule of thumb, you should keep your claude.md file under 200 lines to avoid it.

“And this is exactly the problem we're trying to solve called context rot. So, this is the inability for AI and LLM models to actually recall 100% of the information that we've loaded in as we increase the amount of context that's loaded in the first place. So, as a general rule of thumb, keep your Claude.md under 200 lines.”

You should only move to Level 3 from Level 2 if two conditions are met: (1) you've been using Claude Code for more than a month with memory files grown past a handful, and (2) you've ever typed a question where you know the answer is in your notes but Claude couldn't find it.

“So only add to level two if these two things are true. One, you've been using Claude code for more than a month and your memory files have grown past a handful of files. And two, if you've ever typed a question to Claude where you know the answer is somewhere in your notes, but Claude couldn't actually find it, then it might be a sign to move on to level three.”

At Level 2, when keyword search is used on a large collection of memory files that summarize key topics rather than storing verbatim text, keyword search becomes ineffective as the collection scales, requiring a move to semantic search systems like Level 3.

“as you start to scale this, when you're running it for real for months across multiple clients or projects, the structure eventually stops scaling. So, you end up with a giant general.md file that can't be read efficiently. And because it's summarizing key topics and not taking words verbatim, the keyword search starts falling apart as well.”

LightRAG is an enterprise-grade knowledge graph designed for heavier entity extraction and dual-level retrieval, but it's completely overkill for 99% of business owners and you're better off either staying with Level 4 or using Level 5 alternatives like Karpathy or Recall.

“There's also a quick mention for a heavyweight alternative, which is similar like knowledge graph style, called LightRAG. So, LightRAG is an enterprise-grade knowledge graph, and it's designed for heavier entity extraction, dual level retrieval, the full research grade thing, basically. But, it's completely overkill for 99% of business owners. You don't need the setup headache. You don't need the massive database in the background for that, and you're better off either going back to level four, or if you're looking for a visual knowledge base, then you can install Carpet these or use Recall.”

Levels 1, 2, and 3 can stack together with no issues and have fairly similar folder structures; users can ask Claude Code to integrate them

“the best part about all the things that we covered today is a lot of them do stack together. So you can run levels one, two, and three together with no issues. And actually the folder structure is fairly similar, and you can ask Claude code to actually integrate those yourself.”

Claude Mem stores everything in the background (unlike MemSearch), which has the advantage of automatic capture but the disadvantage of non-readable storage, whereas MemSearch keeps everything in plain markdown so you can go back and read it directly.

“And the other major difference from Claude Mem and Mem search is Claude Mem is actually going to store everything in the background, whereas Mem search is going to keep everything in plain markdown, so you can actually go back and read it.”

Speaker recommends Recall if you're non-technical and mainly want to organize articles, videos, and podcasts, but if you want maximum control and still want to build out that Wikipedia, then Karpathy's LLM Wiki will definitely beat it.

“So, my take is if you're non-technical, and you mainly want to organize articles, videos, and podcasts, then Recall is great. But, if you want maximum control, and you still want to build out that Wikipedia, then the LLM Wiki will definitely beat it there.”

For someone literally just starting with Claude Code, don't worry about memory systems and just turn on Level 1 (claude.md and memory.md files) which takes 10 minutes and is a massive improvement on inbuilt memory.

“So, if you are literally just starting, then don't worry about these memory systems. Turn on level one by just utilizing your claw.md and your memory.md files in the right way. It's going to take you 10 minutes, and it's going to be a massive improvement on the inbuilt memory inside claw.”

Skip Level 6 (Open Brain) if you're still living mostly inside Claude Code on one machine; it's only necessary if you work across multiple AI tools.

“But if you're still living mostly inside code on one machine, then it's probably worth skipping this part.”