diff --git a/README.md b/README.md
index 451ec46..b788738 100644
--- a/README.md
+++ b/README.md
@@ -58,6 +58,7 @@ This is an early work-in-progress. Follow [me on twitter](https://x.com/klntsky)
   - [x] `[# comments]`
   - [ ] `[:status some-status]` - to show during prompt evaluation
   - [ ] `[:call ffi-function :param1=foo :param2=bar]`
+  - [ ] syntax for ignoring `$` output
 - [ ] Implement an evaluator
   - [x] meta-prompting
   - [x] conditionals
@@ -70,7 +71,7 @@ This is an early work-in-progress. Follow [me on twitter](https://x.com/klntsky)
 - [ ] Runtime system
   - [x] Support variable definition at runtime
   - [x] dynamic model switching (via `MODEL` variable - [example](./examples/model-change.metaprompt))
-  - [ ] Multiple chat instances and ability to switch between them, to distribute data between chat contexts. E.g. `[chat1$ the object is the moon][chat2$ the object is the sun][chat1$ what is the object?]`
+  - [x] Multiple chat instances and ability to switch between them, to distribute data between chat contexts. E.g. `[chat1$ the object is the moon][chat1$ what is the object?]` [(example)](./examples/chat-history.metaprompt)
   - [ ] message role system (system, user)
   - [ ] exceptions
     - [ ] throwing exceptions
diff --git a/examples/chat-history.metaprompt b/examples/chat-history.metaprompt
new file mode 100644
index 0000000..ed55814
--- /dev/null
+++ b/examples/chat-history.metaprompt
@@ -0,0 +1,19 @@
+[:_=[chat1$ the $OBJECT is a car. remember this]]
+[:_=[chat2$ the $OBJECT is an apple. remember this]]
+[:_=[chat1$ the $ACTION is "ride". remember this]]
+[:_=[chat2$ the $ACTION is "eat". remember this]]
+[:question=
+  Combine the $ACTION and the $OBJECT
+  into a single phrase. Give me just the phrase,
+  no other output
+]
+chat1: [chat1$ [:question]]
+chat2: [chat2$ [:question]]
+
+Chat history for chat1:
+
+[:chat1]
+
+Chat history for chat2:
+
+[:chat2]
diff --git a/python/src/eval.py b/python/src/eval.py
index f391607..0d51ae2 100644
--- a/python/src/eval.py
+++ b/python/src/eval.py
@@ -11,8 +11,18 @@
 """
 
 
+def serialize_chat_history(chat_history):
+    res = ""
+    for chat_item in chat_history:
+        role = chat_item["role"]
+        content = chat_item["content"]
+        res += f"[{role}]: {content}\n"
+    return res
+
+
 async def eval_ast(ast, config, runtime):
     env = Env(**config.parameters)
+    chats = dict()
     default_model = config.providers.get_default_model()
     if default_model is not None:
         env.set("MODEL", default_model.strip())
@@ -47,9 +57,11 @@ def get_current_model_provider():
             raise ValueError(f"Model not available: {model_name}")
         return provider
 
-    async def stream_invoke(prompt: str) -> AsyncGenerator[str, None]:
+    async def stream_invoke(
+        prompt: str, history=[]
+    ) -> AsyncGenerator[str, None]:
         provider = get_current_model_provider()
-        async for chunk in provider.ainvoke(prompt, "user"):
+        async for chunk in provider.ainvoke(prompt, "user", history):
             yield chunk
 
     async def invoke(self, prompt: str) -> str:
@@ -59,7 +71,7 @@ async def invoke(self, prompt: str) -> str:
         return res
 
     async def _eval_ast(ast):
-        nonlocal env, runtime
+        nonlocal env, runtime, chats
         if isinstance(ast, list):
             # TODO: is this case needed?
             async for chunk in _eval_exprs(ast):
@@ -91,13 +103,20 @@ async def _eval_ast(ast):
                 evaluated_parameters[parameter] = await _collect_exprs(
                     parameters[parameter]
                 )
+            # save parent state in a closure
             old_env = env
+            old_chats = chats
+            # prepare new state
             if "MODEL" not in evaluated_parameters:
                 evaluated_parameters["MODEL"] = get_model()
             env = Env(evaluated_parameters)
+            chats = {}
+            # recurse
             async for chunk in _eval_ast(loaded_ast):
                 yield chunk
+            # restore parent state
             env = old_env
+            chats = old_chats
         elif ast["type"] == "assign":
             var_name = ast["name"]
             value = (await _collect_exprs(ast["exprs"])).strip()
@@ -105,14 +124,36 @@ async def _eval_ast(ast):
                 runtime.set_status(value)
             env.set(var_name, value)
         elif ast["type"] == "meta":
+            # Load chat history
+            chat_id = ast["chat"] if "chat" in ast else None
+            if chat_id is not None:
+                if chat_id not in chats:
+                    chats[chat_id] = []
+            # evaluate the prompt
             chunks = []
-            chat_id = ast["chat"]
             for expr in ast["exprs"]:
                 async for chunk in _eval_ast(expr):
                     chunks.append(chunk)
             prompt = "".join(chunks)
-            async for chunk in stream_invoke(prompt):
+            # collect the assistant response
+            assistant_response = ""
+            async for chunk in stream_invoke(
+                prompt, chats[chat_id] if chat_id in chats else []
+            ):
+                assistant_response += chunk
                 yield chunk
+            # update chat history
+            if chat_id is not None:
+                chats[chat_id].append(
+                    {
+                        "role": "user",  # TODO: use current role
+                        "content": prompt,
+                    }
+                )
+                chats[chat_id].append(
+                    {"role": "assistant", "content": assistant_response}
+                )
+                env.set(chat_id, serialize_chat_history(chats[chat_id]))
         elif ast["type"] == "exprs":
             for expr in ast["exprs"]:
                 async for chunk in _eval_ast(expr):
diff --git a/python/src/providers/interactive.py b/python/src/providers/interactive.py
index f95829e..3da7e60 100644
--- a/python/src/providers/interactive.py
+++ b/python/src/providers/interactive.py
@@ -26,7 +26,8 @@ def __init__(self, api_key: str = None):
     async def ainvoke(
         self,
         prompt: str,
-        role: str = "user"
+        role: str = "user",
+        history = [] # TODO: make interactive provider respect history?
     ) -> AsyncGenerator[str, None]:
         """Asynchronously invoke the OpenAI API and yield results in chunks.
 
diff --git a/python/src/providers/openai.py b/python/src/providers/openai.py
index bb8a164..a27ef12 100644
--- a/python/src/providers/openai.py
+++ b/python/src/providers/openai.py
@@ -3,7 +3,7 @@
 
 import openai
 import os
-from typing import AsyncGenerator
+from typing import AsyncGenerator, List
 
 
 class OpenAIProvider(ProviderConfig):
@@ -44,7 +44,7 @@ def __init__(self, api_key: str = None, model: str = "gpt-4"):
                 "API key is required for OpenAI API. Specify OPENAI_API_KEY environment variable or provide an api_key argument"
             )
 
-    async def ainvoke(self, prompt: str, role: str) -> AsyncGenerator[str, None]:
+    async def ainvoke(self, prompt: str, role: str, history: List[{ "role": str, "content": str }] = []) -> AsyncGenerator[str, None]:
         """Asynchronously invoke the OpenAI API and yield results in chunks.
 
         Args:
@@ -58,7 +58,7 @@ async def ainvoke(self, prompt: str, role: str) -> AsyncGenerator[str, None]:
         # TODO: use system message role for IF_PROMPT
         stream = await client.chat.completions.create(
             model="gpt-4",
-            messages=[{"role": role, "content": prompt}],
+            messages=history + [{"role": role, "content": prompt}],
             stream=True,
         )