Transcript
00:00:00Claude just made the same mistake for the third time, not because the model is dumb,
00:00:04because it has no memory of your project. It doesn't remember the decision from last week,
00:00:09it doesn't remember why you moved to GraphQL, and it doesn't remember your weird conversations.
00:00:14This is MemPallas, and it's blown up with over 52,000 stars on GitHub. It's said to give us
00:00:20local, lossless, long-term memory for our AI tools. I'll show you what MemPallas is,
00:00:26how it works, and the key differences to other tools we already have.
00:00:35Now, the short version of all this is MemPallas stores our conversations word for word, runs locally,
00:00:43and has Claude code hooks and supports MCP tools, and it doesn't need API keys for the core memory flow.
00:00:49Now, the main point I should touch on here is someone's bound to say,
00:00:52we already have context mode. Yes, true, we do. Context mode is good for keeping one long coding
00:00:59session from blowing up your context window. MemPallas is different. It gives your AI permanent
00:01:06memory access completely across separate chats and even days. They actually work great together,
00:01:12so you can actually run both of them side by side. MemPallas is super straightforward,
00:01:17so let me show you how to set this up and how this would work in the real world. If you enjoy coding
00:01:22tools that speed up your workflow, be sure to subscribe. We have videos coming out all the time.
00:01:27Let's just start with the demo because this is where a lot of memory tools just fall apart.
00:01:30I'm going to run UV tool install MemPallas. The repo currently recommends installing it with UV
00:01:37to avoid dependency conflicts, though PIP still works if that's your setup. Then I'm going to initialize a
00:01:43palace for my project just by running MemPallas in it within a project. You're creating a local memory
00:01:51database for the things your AI shouldn't forget. Now you mine the project files, notes, and old
00:01:58clod code sessions. I can do this by saying MemPallas mine. Then I'm going to MemPallas mine the clod
00:02:06conversations from the past. And this is key here because real project context is never in one clean
00:02:11place. It's in commits, docs, chats, notes, random markdown files, and you barely remember making half
00:02:18of these. Now I can ask MemPallas search in a question, like why did we switch to GraphQL? And instead
00:02:25of guessing from today's prompt, it pulls the old discussions where that decision happened. That's the
00:02:32first really big contrast here. Your AI doesn't need a bigger prompt, it just needs better memory.
00:02:37When we run MemPallas mine, this is the step that teaches it everything. It goes into the folder we tell
00:02:44it, it reads every file it finds, and it breaks it into smaller pieces. It then offloads that and saves
00:02:50everything verbatim into your memory palace so then we can search for it. Now, what does MemPallas
00:02:55actually do it, right? The demo worked, we saw how it works. It uses a memory palace metaphor,
00:03:00projects or people becoming the actual infrastructure, related to facts and events
00:03:06live in different rooms. The original source text lives in these drawers, but don't get distracted by
00:03:12that metaphor. The important part is the original content stays intact. Most memory systems do the
00:03:18obvious thing first. They take the messy conversation and ask an LLM to turn it into clean facts. That sounds
00:03:25smart, but it has a big problem. If the summary drops a weird constraint, an edge case, or a reason behind
00:03:31a decision, that detail is gone from memory. MemPallas takes the opposite bet. It keeps the original text
00:03:38word for word and then builds a compact index above it, so the AI can find the right area without
00:03:44throwing away what's inside. Under it all, it is pretty dev friendly. ChromaDB on disk for retrieval,
00:03:51SQLite for the knowledge graph. MCP integration so agents can actually use the memory and claw to code
00:03:57hooks so this can fit into a real coding workflow instead of sitting off the side. MemZero and Zapp
00:04:03are often more productized, more SDK focused, and better if you're building memory into an app or
00:04:09product. MemPallas feels more like a tool for devs who want their coding agents to remember the actual work
00:04:16history locally. The trade-off here is clear. If you want a hosted product, other tools are going
00:04:21to be better. If you want a local memory system that keeps your original source, MemPallas has the edge
00:04:27here. Now, the code-clawed integration is where it starts to feel more practical. The plugin can start the
00:04:33MCP server, expose the tools, teach Clawed the memory protocol, and search the palace before answering
00:04:39questions about past work. The detail sounds small, but it's not. There's a huge difference between
00:04:45I have a database somewhere and my coding agent knows when to look in memory before it answers.
00:04:50Here's what devs seem to like. First, exact recall. If you said something weird but important three
00:04:57weeks ago, it can retrieve the original wording instead of a cleaned up summary. Then, there's cost and
00:05:03privacy. The official docs say the content stays local using ChromaDB and SQLite on disk with no API
00:05:11key required for core flow. Then is the low startup context. You don't need to paste 40,000 tokens of
00:05:18a project lore every time. It's going to run this. You load enough pointers for the agent to know where
00:05:24to look. The temporal knowledge graph is also a big deal here because software decisions, they do expire.
00:05:31We use REST API. That might have been true a couple months ago, and then it might be wrong last month.
00:05:38The normal fact databases can actually miss that. A timeline-aware memory system has a better shot
00:05:44at understanding what changed and when. That is the second big contrast. Memory is not just about facts,
00:05:50it's about time. The palace hierarchy is clever, but it's not perfect by any means. For some projects,
00:05:57plain ChromaDB semantic search over your transcripts might give you most of your value with fewer
00:06:02concepts. There is also set up in the maintenance, so local databases are great until you need backups,
00:06:08cleanup, migration, or a sync. And this is not a full manager to manage your memory platform. If you
00:06:15want admin controls, dashboards, permissioning, this tool is not there yet. One practical warning here,
00:06:21because this project went viral, there are lookalike domains. The readme says the official sources are
00:06:28the GitHub repo and the Python packaging and the docs site. So do not install from random Mem Palace
00:06:34websites you find and search. That is not a small warning. This is a tool that touches your project
00:06:39history, so treat it like part of your dev environment. If you are using cloud code,
00:06:44cursor, or local agents, this makes sense. Or if you care about local first tooling and
00:06:48you just hate paying for subscriptions, just so your AI remembers what you already told it.
00:06:53I would just steer clear of this if you want a zero-config hosted service,
00:06:57or if your projects are short enough that a readme and a good prompt already solve the problem.
00:07:02If you enjoy coding tools like this, be sure to subscribe to the BetterStack channel. We'll see you in another video.
Community Posts
No posts yet. Be the first to write about this video!
Write about this video