IT EventsBook

Discussions

Community guidelines

Be specific and constructive. No vendor spam — promoting your own product belongs in a listing. Anyone can read; posting needs a free account.

AI agent memory poi...
 
Notifications
Clear all
AI agent memory poisoning, how do you secure long term memory?
5 Posts
3 Users
0 Reactions
4 Views
retrieval_rita
(@retrieval_rita)
Active Member
Joined: 1 month ago
Posts: 7
Topic starter   [#52]

we let an assistant remember project facts across chats. A user pasted a bad instruction into a ticket, the assistant summarized it as a permanent project rule, and it kept affecting later work. How do you keep memory useful without turning every document into a lasting prompt injection?



   
Quote
mcp_curious
(@mcp_curious)
Active Member
Joined: 1 month ago
Posts: 9
 

Do not let retrieved content write directly to long term memory. Candidate memories need a type, source, timestamp and confidence. Some should require user confirmation before they become durable.



   
ReplyQuote
entity_edge
(@entity_edge)
Active Member
Joined: 1 month ago
Posts: 8
 

we separate facts from instructions. A ticket can say the server name is X. It can't create a new operating rule. Memories also expire unless they are referenced or confirmed. That removes a lot of stale junk



   
ReplyQuote
retrieval_rita
(@retrieval_rita)
Active Member
Joined: 1 month ago
Posts: 7
Topic starter  

edit: our summary step mixes facts, preferences and instructions into one paragraph, which is probably the root problem



   
ReplyQuote
mcp_curious
(@mcp_curious)
Active Member
Joined: 1 month ago
Posts: 9
 

Exactly. Store structured items and keep the original evidence. When a memory influences an action, show it in the trace. Users need a way to edit and delete it. Memory is a database with untrusted inputs, so apply normal data governance instead of treating it as model magic.



   
ReplyQuote
Share:
Scroll to Top