feat: add relevant memories to retain so it's smarter (#24)

* fix forgetting vector or triple stores in ephemeral mode

* add relevant memories to retain

* fix tests

* remove debug

memonto/core/forget.py

@@ -14,14 +14,14 @@ def _forget(
 ) -> None:
     if ephemeral:
         data.remove((None, None, None))
+    else:
+        try:
+            if vector_store:
+                vector_store.delete(id)
 
-    try:
-        if vector_store:
-            vector_store.delete(id)
-
-        if triple_store:
-            triple_store.delete_all(id)
-    except ValueError as e:
-        logger.warning(e)
-    except Exception as e:
-        logger.error(e)
+            if triple_store:
+                triple_store.delete_all(id)
+        except ValueError as e:
+            logger.warning(e)
+        except Exception as e:
+            logger.error(e)

memonto/core/retain.py

@@ -145,6 +145,24 @@ def update_memory(
         return str(updated_memory)
 
 
+def find_relevant_memories(
+    data: Graph,
+    vector_store: VectorStoreModel,
+    message: str,
+    id: str,
+    ephemeral: bool,
+) -> str:
+    relevant_memory = ""
+
+    if ephemeral:
+        relevant_memory = str(data.serialize(format="turtle"))
+    else:
+        relevant_memory = str(vector_store.search(message=message, id=id, k=3))
+
+    logger.debug(f"relevant_memory\n{relevant_memory}\n")
+    return relevant_memory
+
+
 def save_memory(
     ontology: Graph,
     namespaces: dict[str, Namespace],
@@ -158,12 +176,21 @@ def save_memory(
     str_ontology: str,
     updated_memory: str,
 ) -> None:
+    relevant_memory = find_relevant_memories(
+        data=data,
+        vector_store=vector_store,
+        message=message,
+        id=id,
+        ephemeral=ephemeral,
+    )
+
     script = llm.prompt(
         prompt_name="commit_to_memory",
         temperature=0.2,
         ontology=str_ontology,
         user_message=message,
         updated_memory=updated_memory,
+        relevant_memory=relevant_memory,
     )
 
     logger.debug(f"Retain Script\n{script}\n")

memonto/prompts/commit_to_memory.prompt

@@ -10,14 +10,20 @@ And this user message:
 ${user_message}
 ```
 
-And these removed triples:
+And these removed memories:
 ```
 ${updated_memory}
 ```
 
+And these relevant memories:
+```
+${relevant_memory}
+```
+
 Analyze the user message to find AS MUCH new information AS POSSIBLE that could fit onto the above ontology while adhering to these rules:
-- First find all the new information in the user message that maps onto BOTH the above ontology and ESPECIALLY the removed triples.
-- Second apply the existing namespaces to the extracted information.
+- First find all the new information in the user message that maps onto BOTH the above ontology and ESPECIALLY the removed memories.
+- Second check if the relevant memories can help extract even more information from the user message that maps onto the ontology or removed memories.
+- Third apply only the existing namespaces to the extracted information.
 - Finally create the script that will add the extracted information to an rdflib graph called `data`.
 - NEVER generate code that initializes new graphs, namespaces, classes, properties, etc.
 - GENERATE Python code to add the triples with the relevant information assuming rdflib Graph `data` and the newly added namespaces already exists.

tests/core/test_retain.py

@@ -66,6 +66,7 @@ def test_commit_memory(
         ontology=ANY,
         user_message=user_query,
         updated_memory="",
+        relevant_memory=ANY,
     )
 
     assert mock_llm.prompt.call_count == 1
@@ -103,6 +104,7 @@ def test_commit_memory_with_exception(
         ontology=ANY,
         user_message=user_query,
         updated_memory="",
+        relevant_memory=ANY,
     )
 
     ctmeh_prompt = call(
@@ -146,6 +148,7 @@ def test_commit_memory_auto_expand(
         ontology=ANY,
         user_message=user_query,
         updated_memory="",
+        relevant_memory=ANY,
     )
 
     eo_prompt = call(