Unverified50% confidenceSolutionExact time
Stop appending raw JSON payloads to working memory, implement aggressive pruning filters, and extract only verified entity fields before updating the conversation context.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
Unverified代理层若在转换中改变JSON序列化顺序、注入动态时间戳或调整消息结构，会导致前缀不一致使缓存彻底失效70% similarUnverifiedMalicious instructions from indirect prompt injection are hidden in external content that the agent reads (web pages, PDFs, database fields, JSON returned by tools), and the traditional approach of filtering at the user input is ineffective against this; the effective interception point is to check before tool results enter the context.69% similarUnverifiedAdding JSON tags to retrieved memory rows does not solve prompt injection; context should be placed in the data portion of the request rather than in system instructions that carry higher priority67% similarUnverifiedJSONL格式支持追加写入，中断前的消息可安全落盘，相比JSON数组存储更适合AI代理会话记录场景66% similarUnverifiedThe replay functionality is implemented by recording a JSON log of operation sequences, which can be directly re-executed without re-invoking the large model to save inference costs.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1044041API
curl https://kongchang.com/api/v1/knowledge/claims/1044041MCP
get_claim(id=1044041)