diff --git a/src/interface.jl b/src/interface.jl
index 6434795..f46aaaa 100755
--- a/src/interface.jl
+++ b/src/interface.jl
@@ -314,15 +314,16 @@ function planner_mistral_openorca(a::agentReflex)
   prompt = replace(prompt, "{toolnames}" => toolnames)
   prompt = replace(prompt, "{tools}" => "You have access to the following tools:\n$toollines")
   
+
+  # prepare contex
   context = 
-    """
-    {earlierConversation}
-    {env state}
-    {longterm memory}
-    """
-  context = replace(context, "{earlierConversation}" => "My earlier talk with the user:\n$(a.earlierConversation)")
-  context = replace(context, "{env state}" => "")
-  context = replace(context, "{longterm memory}" => "")
+  """
+  My earlier talk with the user:
+  $(a.earlierConversation)
+
+  My earlier experience
+  $(experience(a.memory[:longterm]))
+  """
 
   prompt = replace(prompt, "{context}" => context)
 
@@ -379,7 +380,7 @@ function actor_mistral_openorca(a::agentReflex)
   prompt = replace(prompt, "{thinkingFormat}" => a.thinkingFormat[:actor])
   prompt = replace(prompt, "{step}" => a.step)
   
-  s = shortMemoryToString(a.memory[:shortterm], ["user:", "Plan 1:"])
+  s = dictToString(a.memory[:shortterm], ["user:", "Plan 1:"])
   prompt = replace(prompt, "{shorttermMemory}" => s)
 
   toolnames = ""
@@ -457,7 +458,7 @@ function conversation(a::T, usermsg::String) where {T<:agent}
     thinkingmode = chooseThinkingMode(a, usermsg)
     @show thinkingmode
     if thinkingmode == :no_thinking 
-      a.context = conversationSummary(a)  #TODO should be long conversation before use summary because it leaves out details
+      a.context = conversationSummary(a)
       _ = addNewMessage(a, "user", usermsg)
       prompt = generatePrompt_mistral_openorca(a, usermsg, thinkingmode)
       @show prompt
@@ -600,12 +601,13 @@ function conversation(a::agentReflex, usermsg::String; attemptlimit::Int=3)
   a.attemptlimit = attemptlimit
   respond = nothing
 
+  a.earlierConversation = conversationSummary(a)
+
   # determine thinking mode
   a.thinkingmode = chooseThinkingMode(a, usermsg)
   @show a.thinkingmode
 
   if a.thinkingmode == :no_thinking
-    a.earlierConversation = conversationSummary(a)  #TODO should be long conversation before use summary because it leaves out details
     _ = addNewMessage(a, "user", usermsg)
     prompt = chat_mistral_openorca(a, usermsg)
     println("")
@@ -628,12 +630,11 @@ function work(a::agentReflex, usermsg::String)
   respond = nothing
 
   if a.thinkingmode == :new_thinking
-    a.earlierConversation = conversationSummary(a)
     _ = addNewMessage(a, "user", usermsg)
     a.memory[:shortterm]["user:"] = usermsg
     a.memory[:log]["user:"] = usermsg
     a.newplan = true
-  elseif a.thinkingmode == :continue_thinking #TODO
+  elseif a.thinkingmode == :continue_thinking
     println("continue_thinking!!")
     _ = addNewMessage(a, "user", usermsg)
     a.memory[:shortterm]["Obs $(a.step):"] = usermsg
@@ -710,7 +711,6 @@ function work(a::agentReflex, usermsg::String)
 
           a.memory[:shortterm] = keepOnlyKeys(a.memory[:shortterm], ["user:"])
 
-          #TODO add lesson and context into longterm memory
           headerToDetect = ["Lesson:", "Context:", ]
           headers = detectCharacters(lessonwithcontext, headerToDetect)
           chunkedtext = chunktext(lessonwithcontext, headers)
@@ -723,7 +723,7 @@ function work(a::agentReflex, usermsg::String)
         error("undefied condition, actorstate $actorstate $(@__LINE__)")
         break
       end
-    else  #TODO attempt limit reached, force AI to answer
+    else
       error("attempt limit reach")
       break
     end
@@ -854,7 +854,7 @@ function actor(a::agentReflex)
         a.memory[:shortterm]["Obs $(a.step):"] = toolresult
         a.memory[:log]["Obs $(a.step):"] = toolresult
       end
-    else  #TODO finish all steps
+    else
       actorState = "all steps done"
       msgToUser = nothing
       break
@@ -999,9 +999,9 @@ julia> shorttermMemory = OrderedDict{String, Any}(
   "Obs 2:"      => "No info available for your search query.")
 julia> report = analyze(agent, shorttermMemory)
 ```
-""" #WORKING analyze sometime result in empty string ""
+"""
 function analyze(a)
-  shorttermMemory = shortMemoryToString(a.memory[:shortterm], ["user:"])
+  shorttermMemory = dictToString(a.memory[:shortterm], ["user:"])
   prompt =
   """
     <|im_start|>system
@@ -1110,7 +1110,7 @@ julia> report = formulateUserRespond(agent, shorttermMemory)
 function formulateUserRespond(a)
   stimulus = a.memory[:shortterm]["user:"]
 
-  work = shortMemoryToString(a.memory[:shortterm], ["user:"])
+  work = dictToString(a.memory[:shortterm], ["user:"])
 
   prompt =
   """
@@ -1164,10 +1164,10 @@ julia> shorttermMemory = OrderedDict{String, Any}(
 julia> decision = goNogo(agent)
 "Yes"
 ```
-""" #BUG sometime AI ready to formulate respond before all step are completed
+"""
 function goNogo(a)
   stimulus = a.memory[:shortterm]["user:"]
-  work = shortMemoryToString(a.memory[:shortterm], ["user:"])
+  work = dictToString(a.memory[:shortterm], ["user:"])
 
   # prompt =
   # """
diff --git a/src/utils.jl b/src/utils.jl
index 2a3084e..c59f499 100644
--- a/src/utils.jl
+++ b/src/utils.jl
@@ -3,7 +3,7 @@ module utils
 export makeSummary, sendReceivePrompt, chunktext, extractStepFromPlan, checkTotalStepInPlan,
         detectCharacters, findDetectedCharacter, extract_number, toolNameBeingCalled,
         chooseThinkingMode, conversationSummary, checkReasonableness, replaceHeaders,
-        addShortMem!, splittext, shortMemoryToString, removeHeaders, keepOnlyKeys
+        addShortMem!, splittext, dictToString, removeHeaders, keepOnlyKeys, experience
 
 using UUIDs, Dates, DataStructures
 using CommUtils, GeneralUtils
@@ -431,7 +431,7 @@ function conversationSummary(a::T) where {T<:agent}
   conversation = ""
   summary = "nothing"
   if length(a.messages)!= 0
-    for msg in a.messages
+    for msg in a.messages[1:end-1]
       role = msg[:role]
       content = msg[:content]
 
@@ -638,7 +638,7 @@ julia> shortMemory = OrderedDict(
                               )
 
 julia> headers = ["user:"]
-julia> shortMemoryToString(shortMemory, headers)
+julia> dictToString(shortMemory, headers)
 "Thought 1: I like it.\nAct 1: chatbox\nActinput 1: I get this one.\n"
 ```
 """
@@ -660,6 +660,24 @@ function shortMemoryToString(shortMemory::OrderedDict,
 end
 
 
+function dictToString(dict::T, 
+          skiplist::Union{Array{String}, Array{Symbol}}) where {T<:AbstractDict}
+  s = ""
+  for (k, v) in dict
+    if k ∉ skiplist
+      s1 = "$k $v"
+      s *= s1
+
+      # ensure a newline seperate each sentences
+      if s[end] != "\n"
+        s *= "\n"
+      end
+    end
+  end
+  return s
+end
+
+
 """ Remove headers of specific step from memory.
 
 Args:
@@ -764,6 +782,57 @@ function keepOnlyKeys(dict::T1, keys::T2) where {T1<:AbstractDict, T2<:AbstractV
   return newdict
 end
 
+
+""" Convert experience dict into 1 string for LLM to use.
+
+Args:
+  dict = a dictionary contain past experience
+
+Return:
+  An experience in 1 string without context keys.
+
+# Example
+```jldoctest
+julia> dict = OrderedDict{String, Any}(
+                " This lesson can be applied to various situations => " Gathering accurate and relevant information about the user's preferences, budget, and event details is crucial for providing personalized recommendations.\n"
+                )
+julia> experience(dict)
+
+```
+"""
+function experience(dict::T) where {T<:AbstractDict}
+  s = ""
+  for (k, v) in dict
+    s *= v
+  end
+
+  return s
+end
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+