been sitting with MemSync for a couple days now and the part i keep circling isnt the feature itself its the infrastructure underneath it....
heres the mechanic.MemSync extracts memories from conversations, documents, websites, social profiles all using TEE-verified LLM calls. so its not just storing what you told it. the extraction process itself is cryptographicaly attested. then memories get classified as either semantic lasting facts like "software engineer at google" or episodic time-bound things like "currently working on an ios app." the distinction matters because the system treats them diferently in retrieval....
not a database.a living profile.
and then theres the semantic search layer, which i think is the part most people dont think about until they need it. you query your memory using natural language with embedding-based similarity. you dont have to remember exactly what you told it it finds the relevant context for y0u....
i actualy find this reassuring in a narrow way. the entire memory pipeline runs on verifiable infrastructure extraction,classification,profile generation, maintenance. that means the AI building your memory profile is itself verifiable, not just the storage....
but i wont pretend verifiable memory extraction is the same as accurate memory extraction.the LLM deciding what counts as a semantic fact versus an episodic event could still miss-clasify things in ways that compound over time....
about a year ago i started using a popular AI memory tool and realised after smething like three months that it had been storing surface-level observations rather than anything actualy usefull. the retrieval was fast but the memory was shallow. made me think harder about what extraction quality realy means....
what i still cant resolve is how MemSync handles conflicting memories if an episodic fact becomes outdated and a new one contradicts it, does the system overwrite,flag the conflict, or carry both versions forward??
