Weft is one of the first systems I built when I started working with agentic coders, and I am pleased to announce that after months, I think it’s finally ready for release.
[!Tip] The TLDR - Don’t Have Time To Read? Here’s The Key Points
- Weft is a memory system for your AI
- Weft is MIT Licensed and Free out of the box
- Weft Is Local - It Runs in a Docker container on your computer
- I designed Weft for code. I do not use a personal assistant (OpenClaw,Hermes) so I am not sure how it will act in that system. If you try it, please let me know!
- Weft is not designed to remember everything about you across all context. It is designed to give your agent the context it needs for the task at hand and nothing more.
If that’s still too much for you, point your robot at https://weft.mediumroast.dev. Ask them what they think.
50 First Chats
AI agents are smart. They can understand a codebase, diagnose a problem, and implement a fix..as long as they don’t run out of context.
Every new chat starts from 0.
Unless you install a dedicated memory system you have two choices:
- Store important information in “flat files” like Agents.md on your computer and then always remember to prune them when information becomes outdated
- Use the built in “memory” system in chat and risk it thinking that an offhand reference you made months ago is gospel.
Both of these options have the same issue: They don’t know what’s relevant to you right now so they either load a ton of stuff they don’t need, or make an assumption and start insisting that the idea you’re talking about would work PERFECTLY on that $200,000 AI rig you asked about 3 months ago and your agent decided you owned.
I Think I have something better.
My Favorite Weft Features
Weft Gives Agents Persistence
The core loop of Weft is /handoff and /prime (I have these as skills to make them more natural).
When you’re done with a session, tell your agent to write a Handoff for itself. When you start a new session minutes, days, or weeks later, tell it to Prime. The agent gets a high level summary of where you left off, key decisions, and next steps.
Have something that’s really important? Pin it to load the memory every session or save an anti-pattern if you want to keep your agent from defaulting to a path you don’t like.
Weft Lets Your Agents (And Skills!) Find Key Memories
Individual memories and decisions (as well as your handoffs) are stored using Vector Embedding. This makes them searchable similar to how AI search uses RAG (Retrieval Augmented Generation).
If you build skills to be Weft-aware, they can automatically persist findings, look for previous decisions, and even give themselves their own “memory.” I have a skill I use to review my plans for potential weak areas. It stores a ledger of common patterns it sees in my plans so that it looks for those issues specifically.
Weft Is Project-Based But Globally Aware
By default, when you /prime in a folder, it loads that project. Agents looking for information will default to searching only for memories tagged with that project.
But agents aren’t locked to that project. You can say things like “In my other project we tackled this using X” and Weft can look for memories there instead. If something is super important across all projects, you can save the memories as global (visible to all projects).
Weft Respects Your Context Budget
Your agent only has a limited amount of context to work before you either need to compact the session or start a new one. If Weft loaded everything in, you wouldn’t get a lot of usable development time.
Weft’s Prime has a strict token limit and loads progressively by default, giving your agent just the context it needs, and not a ton of extra “stuff.”
Known Limitations
To date, Weft only runs on two systems. Mine and a trusted friend who helped me stress test some of the core features. I’ve tried making sure that the system loads cleanly, but I don’t know how it will work with your workflow.
I also don’t use any personal assistant harnesses (like OpenClaw). I don’t know how the system will interact with that workflow, though I don’t really see a reason it shouldn’t work.
Benchmark Honesty
For coding, I find Weft to be an invaluable asset, but the benchmarks did highlight a few known limitations.
The problem with how weft saves memory is that it usually doesn’t save stuff until you tell it to. I created “Turn Mode” for people who wanted to use a more assistant memory system, but it’s not perfect. The numbers are ok, but there’s still a few weaknesses. I’m working on it.
If you give Weft a try and run into issues, please submit feedback. I can’t promise to fix everything (I am just one person after all) but this is a system I use everyday. I want it to get better and hope to keep pushing improvements.
What Comes Next
I am hoping that others use Weft and give me feedback so I can improve the product further. I am genuinely proud of what I made and hope that others find it as useful as I do.
I do have a few things on my planned roadmap, however:
- Bringing (Optional) Cloud Functions Into The RC - My personal weft currently runs in Supabase so I can use it across devices. I want to have this be a part of the official release so people can do this if they want.
- Fixing The Issues Surfaced In The Benchmark - Temporal recall is frustrating but super important. I will crack this.
- Allowing Query Fan Out - For complex questions, I want to make it easier for the agent to ask a bunch of related questions (I think this will help with temporal recall). RIght now I plan on investigating AI’s like Jev or Needle to help.
- A Daily Digest That Works - I had this idea that Weft could send me a “status” of everything as a todo. This… does not work currently. I want to make it suck less.
- Shared Memories - This is longer term, but I think it would be cool if two people who have agents working in a shared repo could share memories. This has a TON of security implementations though, so I am taking this slow.
Getting Started With Weft
The easiest way to get started with weft is to point your agent at the Github repo or the Weft site.
There’s instructions there for you to install it if you want, but I tried taking an “agents first” approach.
Weft is Free to use and has an MIT License. This means you can take it, fork it, and customize it however you want. I designed the default install to be local and use only free tools (Docker and FastEmbed) though you might have better results by using a more powerful Embedding API.
I use this as my own system and so I tried making it as robust and functional as possible, but setting up a backup is recommended (there is a weft backup function).
I Highly recommend reviewing the sample agents.md file as well as the example /prime and /handoff skills since this makes the service a bit more seamless.
If you do use Weft, I’d love to hear about your experience (positive and negative). I made Weft in February and decided a bit over a month ago to clean up the repo so that I could share it with others.
I hope you find it as invaluable as I do.