Launched this week
LISA Core
LLM memory using semantic compression for AI conversations
57 followers
LLM memory using semantic compression for AI conversations
57 followers
LISA (Language Intelligence Semantic Anchoring) is a privacy-first browser extension that captures, compresses, and preserves your AI conversations across all major platforms. Using advanced semantic compression technology, LISA achieves 80:1 to 100:1 compression ratios while maintaining complete meaning and context. Your conversations never leave your browser. Zero cloud processing. 100% local.
LISA Core
@amar_dahmani1 the portability angle is clear, but the thing that feels harder over time is knowing what not to carry forward. once someone starts saving a lot of conversations across different models, how do you stop the memory library from becoming technically portable but still too noisy to be useful?
LISA Core
@artem_kosilov The library feature is an actual library feature in which you store your compressed conversations, the extension does the compression locally then it uploads it to the app library, that's the portability, no llm has access to it, you choose which to download and upload. What I usually use it for, is for model colision or switch, I start let's say with sonnet 4.6 dont like the direction the convo is going or unhappy with debugging, I save the conversation switch to Opus and eitheir query the coversation or just ask it to fix the bug sonnet failed to fix, so all is clear on how to use its portability. Basically, the convos are in machine language; no noise just signal :)
Wow this can save a lot of effort and confronting privacy in the start itself is a good move. One question, is the conversation compression model driven or user specified?
LISA Core
@prateek_kumar28 it is not compression per se, it is closer to translation, it translates and disambiguates human language locally it is petty standard how the disambiguation process happens for llms, the extension and app, saves that effort to your model, saving tokens and compute time and so much more downstream, you shoudl give it spin, Id be happy to have feedback
Pretty cool @amar_dahmani1, a real problem and will check out the capabilities. Works on Claude code?
LISA Core
@sayanta_ghosh yes, it does, I built it when I wanted to have Opus teach me how to use Claude code and it worked so I kept it :) Thanks for your interest :)