Pieces for Developers

@nikhil_l Personalized AI is the logical next step with LLMs - Pieces Long-Term Memory makes that happen!

You'd never hire a personal assistant who woke up every morning with amnesia! AI shouldn't be any different.

Thanks for the support and hope you're loving LTM-2!!

@hanna_stechenko2 This is probably the coolest update we have had yet! 🥳

Thanks @steve_caldwell2 , Sorry I was not able to catch what MCP is, can you sherd more light on what MCP means? Happy to answer your question;

@steve_caldwell2 @ialimustufa MCP = Model Context Protocol (https://www.anthropic.com/news/model-context-protocol) a standard how AIs can fetch data, it's good, but not comparable to pieces.

@ialimustufa @henry_rausch Thanks for providing the additional context there Henry. I'm not sure I'd call it "comparable" to Pieces, but it certainly seems like a protocol that Pieces could leverage to quickly connect to many more data sources. It's kind of the hot girl at the AI context dance right now.

@henry_rausch @steve_caldwell2 Great question Steve! We don't currently use MCP, but we're excited about any standardization of data sources. With LTM-2 we're currently focussed more on the contextual data you're explicitly working with, but in the future we are definitely planning on augmenting the context for a query with these external data sources if we feel that it will provide a more useful answer!

@henry_rausch @steve_caldwell2 @ialimustufa Funnily enough I've been playing with MCP recently as a side project. We have a Python SDK so it wouldn't be too hard to implement this yourself for now using our Python SDK and the MCP Python SDK.

@jackross Exactly! LTM-2 is the major upgrade in AI assistants that we have all been waiting for. Thank you for the support Jack!

@sam_parks_at_pieces Workstream Activities are a game-changer for sure 😎

@zongze_x thanks for the great technical question, you are spot on, the issue of memory contamination is complex and a core challenge in designing features like this one. I suspect you can appreciate that writing a full answer here is tough - but would make an excellent topic for a technical article (watch this space). At a high-level, our approach to identifying and minimizing contamination happens at three levels:

• On entry: we are very selective about what is added to the LTM. By analysing where the users focus is and how what they are focusing on currently relates to the big picture of there workstream we can prevent a lot of corruption at source.

• On roll up: when we roll up memories into periodic summaries our agent looks for narratives and themes across workflow elements. When we find contradictions, we resolve them by comparing those narratives to cut out random chatter and keep focus on core tasks.

• At query time: when you interact with your workstream data, through the copilot or the summaries, those interactions are used to infer which aspects are useful/truthful and which are not, which allows us to elevate quality information whilst demoting the noise.
  

Additionally, signals from all of these levels are used to periodically clean contamination from your stored memories. It's a work in progress but I have found the LTM to be much more resistant to context corruption than other solutions out there.

@themisty Hey Krishna! Thank you for checking out the launch! Your paper sounds really interesting, I'm sure a lot of people from my team would enjoy reading it. Where will it be published?

@elliezub Dear Ellie, I am excited to have readers interested already :) it will be on Arxiv, fingers crossed. Still under heavy reviewing lol.

@themisty Sounds great! Looks like we are connected on Linkedin now, so hopefully you will post about it once it's published. Can't wait to read it!

@themisty Can't wait to read your paper Krishna! Thanks for the support as always!!

@elliezub Thanks will definitely ping the team. In the meantime I definitely don't mind if you guys can feel my product nonilion ? I don't mind a solid feedback 🙌