Repository navigation
[RFC] Lineage tracking & permission inheritance on synthesized memory graphs #7887
harishbalasubramanian
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hey team,
Been looking into how Mem0 does memory persistence in multi agent setups, specifically on authorization. Looking at background agents writing derived facts back to the graph in particular.
Direct retrieval with standard metadata filtering is fine for static docs. However, there's a blind spot when background workers synthesize access in different access tiers:
I put tgt an offline benchmark (
sentinel_eval, Apache 2.0) to test direct retrieval vs multi doc memory synthesis. RAG has 0% leaks but synthesized agent memory had leaks in 67% of the tasks since vector search doesn't traverse the permission DAG.Minimal offline test suite here if it helps: https://github.com/harishbalasubramanian/sentinel-eval
How's the team thinking ant this as a graph memory scales?
memory.add()to branch perms before the persistence layerWhat do you all think about permission inheritance and how it should look?
All reactions