← Mnemos

02 / Mnemos - longitudinal evaluation

Evaluating Memory Over Time

A synthetic persona moves through conversations, tools, files, apps, screens, and memory revisions. Evaluation questions arrive only after the evidence they need.

Conversation & tools Filesystem & activity Mnemos memory Evaluation checkpoints Independent event
Sep 02, 2026

Filesystem at this point


    

    

      
Mnemos state at this point

      

    

  


  

  

    FR fact recall
    MU memory updating
    TR temporal reasoning
    CR conflict resolution
    SF selective forgetting
    PU personalization / user modeling
  


  

    

      
How HippoCamp-Chat-S is synthesized

      
First assemble a grounded evidence core. Only then expand it into a realistic, noisy history.

    


    

      

        
        
        
      

      Stage 1 / HippoCamp persona files
      
    


    

      

        Persona workspace / synthesis field

        

          
Maya47 files

          

            
profile.md

            
relocation-notes.pdf

            
calendar-history.json

            
preferences.yaml

          

        


        

          golden file slice
          grounded facts
+ evolving facts
          selected across modalities
        


        
Chat S-12introduces the relocation plan

        
Chat S-31uses a file through a tool

        
Chat S-48updates the earlier fact


        
EEvidence field

        
Qeval question

        
Areference answer


        

        
unrelated
trip chat

        
ambient
tool use

        
old
preference

        
nearby
session

        
file
mention

        
routine
chat


        
      

    


    

      

        Evidence core / stages 1-5
        
Select grounded or evolving facts; synthesize multi-session chat/tool trajectories; create Q, reference A, and gold evidence; validate grounding, chronology, answerability, sufficiency, and no leakage.

        
← reject: return to trajectory + evidence synthesis

      

      

        Distractor expansion / stages 6-7
        
Expand only after review. Distractors cannot answer or contradict gold evidence; deliberate updates must be tagged and reflected in A.

      

      

        Evaluate / stage 8
        
Assemble checkpoints; run a fixed agent/model.