From ea04c0cf30bbf03ec32498e43dd11d9dea867a4f Mon Sep 17 00:00:00 2001
From: Dhanalakshmi M <dhanalakshmi450532@gmail.com>
Date: Tue, 29 Sep 2026 00:23:40 +0530
Subject: [PATCH 1/2] Add support bot memory decisions example

---
 docs/examples.md        | 13 ++++++++
 examples/support_bot.py | 72 +++++++++++++++++++++++++++++++++++++++++
 2 files changed, 85 insertions(+)
 create mode 100644 examples/support_bot.py

diff --git a/docs/examples.md b/docs/examples.md
index ecbca9e..2cd5709 100644
--- a/docs/examples.md
+++ b/docs/examples.md
@@ -1,3 +1,16 @@
 # Examples
 
 See `examples/basic_usage.py` and evaluation datasets in `benchmarks/datasets/`.
+
+## Support Bot with Memory Decisions
+
+See `examples/support_bot.py` for a support-bot example that uses memory decisions to determine how to answer a user query.
+
+The example handles four possible decisions:
+
+- **Replay** — returns the stored response directly.
+- **Restore** — restores relevant context before generating a response.
+- **Verify** — uses a verification tool before generating a response.
+- **None** — makes a fresh LLM call when no useful memory is available.
+
+The LLM and verification tool are stubbed in the example so the focus is on how the memory decisions are handled.
diff --git a/examples/support_bot.py b/examples/support_bot.py
new file mode 100644
index 0000000..6249787
--- /dev/null
+++ b/examples/support_bot.py
@@ -0,0 +1,72 @@
+"""Support bot example using memory decisions."""
+
+from agent_memory import Memory, MemoryAction, MemoryType
+
+
+def call_llm(prompt: str) -> str:
+    """Stub for an LLM call."""
+    return f"LLM response: {prompt}"
+
+
+def verify_with_tool(context: str) -> str:
+    """Stub for an external verification tool."""
+    return f"Verified information: {context}"
+
+
+def support_bot(memory: Memory, query: str) -> str:
+    decision = memory.resolve(query)
+
+    if decision.action == MemoryAction.REPLAY:
+        return decision.response
+
+    if decision.action == MemoryAction.RESTORE:
+        context = memory.format_restore_context(decision)
+        return call_llm(
+            f"Use the following restored context to answer the user:\n"
+            f"{context}\n\nUser: {query}"
+        )
+
+    if decision.action == MemoryAction.VERIFY:
+        context = memory.format_verify_context(decision)
+        verified = verify_with_tool(context)
+        return call_llm(
+            f"Use this verified information to answer the user:\n"
+            f"{verified}\n\nUser: {query}"
+        )
+
+    return call_llm(query)
+
+
+def main() -> None:
+    memory = Memory(persist_dir=".support_bot_demo")
+
+    memory.remember(
+        query="How do I reset my password?",
+        response="Go to Settings → Security → Reset Password and follow the email link.",
+        type=MemoryType.CONVERSATION,
+        tags=["auth", "faq"],
+    )
+
+    memory.remember(
+        query="Current API rate limit",
+        response="1000 requests/minute per API key.",
+        type=MemoryType.FACT,
+        requires_verification=True,
+    )
+
+    queries = [
+        "How do I reset my password?",
+        "Give me a one-sentence explanation of password reset",
+        "What is the API rate limit?",
+        "What's the weather today?",
+    ]
+
+    for query in queries:
+        print(f"\nUser: {query}")
+        print(f"Bot: {support_bot(memory, query)}")
+
+
+if __name__ == "__main__":
+    main()
+
+

From 6571cf91cbf08271be6336a199bfabbb7e447532 Mon Sep 17 00:00:00 2001
From: Dhanalakshmi M <dhanalakshmi450532@gmail.com>
Date: Thu, 1 Oct 2026 00:00:40 +0530
Subject: [PATCH 2/2] Clarify expected support bot responses

---
 examples/support_bot.py | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/examples/support_bot.py b/examples/support_bot.py
index 6249787..23867ba 100644
--- a/examples/support_bot.py
+++ b/examples/support_bot.py
@@ -55,9 +55,16 @@ def main() -> None:
     )
 
     queries = [
+        # Expected action: REPLAY — return the stored response directly.
         "How do I reset my password?",
+
+        # Expected action: RESTORE — restore relevant context, then call the LLM.
         "Give me a one-sentence explanation of password reset",
+
+        # Expected action: VERIFY — verify the stored information, then call the LLM.
         "What is the API rate limit?",
+
+        # Expected action: NONE — no useful memory, so call the LLM directly.
         "What's the weather today?",
     ]