supermemory reposted this
Here's one of the most interesting things we had to decide when building supermemory. and something most people miss. this is why your memory (probably) sucks. and it's not a recall problem, it's not an architecture problem. it's a fundamental problem with "memory" no memory system can correctly identify what's worth remembering. unless you remember everything (which is inefficient and expensive) or just do RAG (which leads to loss of temporal/knowledge updates and no forgetfulness, and worse multi session reasoning) So, we came up with a Hybrid memory model. We _explicitly_ extract and remember what we think is "worth" remembering - this is the stuff that goes into the supermemory graph. Facts build on top of other facts, things are forgotten, things get updated, and knowledge evolves and grows over time. For everything that's not _explicitly_ remembered, we still _implicitly_ remember it, by simply just putting it into our powerful retrieval engine. On search time, you can just set `searchMode` to hybrid - supermemory will traverse the graph and return the answers. And for the blanks where supermemory isn't sure, it returns the chunks as well, with a lower weight. Then, the model can reason over all the context and answer the question perfectly - having the full knowledge of a git-like history as well as raw chunks. Internally, supermemory will learn those chunks for next time, since now we know it's important and may be important when querying next time. supermemory is a self improving system in itself!
