From d2178429176d81393c226ce6e51080a8c27d2319 Mon Sep 17 00:00:00 2001
From: daniel <daniel@ness.family>
Date: Wed, 13 May 2026 11:48:53 -0500
Subject: [PATCH] refactor: remove MEMORY.md from prompt, add AgentCore memory
 instructions

---
 .../app/agent_claw_main/prompt_builder.py     | 39 ++++++++++---------
 1 file changed, 20 insertions(+), 19 deletions(-)

diff --git a/agentclaw/app/agent_claw_main/prompt_builder.py b/agentclaw/app/agent_claw_main/prompt_builder.py
index 782b6b2..823f181 100644
--- a/agentclaw/app/agent_claw_main/prompt_builder.py
+++ b/agentclaw/app/agent_claw_main/prompt_builder.py
@@ -46,25 +46,8 @@ def _get_base_prompt(actor_id: str = '') -> str:
     s3 = boto3.client('s3')
     parts = []
 
-    # Per-user MEMORY.md (falls back to global)
-    memory_key = f'users/{actor_id}/MEMORY.md' if actor_id else 'MEMORY.md'
-    try:
-        obj = s3.get_object(Bucket=bucket, Key=memory_key)
-        content = obj['Body'].read().decode('utf-8')
-        parts.append(f'## MEMORY.md\n{content}')
-        print(f'[prompt_builder] Loaded {memory_key} ({len(content)} bytes)')
-    except ClientError as e:
-        if e.response['Error']['Code'] in ('NoSuchKey', 'AccessDenied') and actor_id:
-            # Fall back to global MEMORY.md
-            try:
-                obj = s3.get_object(Bucket=bucket, Key='MEMORY.md')
-                content = obj['Body'].read().decode('utf-8')
-                parts.append(f'## MEMORY.md\n{content}')
-                print(f'[prompt_builder] Loaded MEMORY.md (fallback, {len(content)} bytes)')
-            except Exception as e2:
-                print(f'[prompt_builder] Failed to load MEMORY.md: {e2}')
-        else:
-            print(f'[prompt_builder] Failed to load {memory_key}: {e}')
+    # MEMORY.md removed — AgentCore Memory handles persistent facts via conversation history.
+    # Long-term memory extraction (retrieval_config + memory strategy) is the right layer for this.
 
     for fname in ['SOUL.md', 'AGENTS.md', 'IDENTITY.md', 'TOOLS.md', 'HEARTBEAT.md']:
         try:
@@ -75,6 +58,24 @@ def _get_base_prompt(actor_id: str = '') -> str:
         except Exception as e:
             print(f'[prompt_builder] Failed to load {fname}: {e}')
 
+    parts.append(
+        '## Memory\n'
+        'Your memory works through two layers:\n\n'
+        '**Conversation history (short-term):** AgentCore automatically loads your full '
+        'conversation history with this user at the start of each session. You have complete '
+        'context of everything discussed previously — no need to ask users to repeat themselves.\n\n'
+        '**Long-term facts (LTM):** Important facts extracted from past conversations are '
+        'retrieved and injected as context automatically. These are things like preferences, '
+        'setup details, names, and recurring topics the user has shared.\n\n'
+        'Guidelines:\n'
+        '- Never ask "what did we discuss last time?" — you already have the history.\n'
+        '- When a user shares something important (job interview, preference, key decision, '
+        'setup change), acknowledge it and trust it will be captured — do not ask if they '
+        'want you to remember it.\n'
+        '- If you notice a fact that seems important but may not be in LTM yet (e.g. a '
+        'deadline, a preference, a name), you may say "I\'ll keep that in mind" — but do '
+        'not ask permission or make a production of it.'
+    )
     parts.append('## Runtime\nRuntime: agent-claw | host=AgentCore | model=bedrock-claude-sonnet | channel=telegram | timezone=America/Chicago')
 
     result = '\n\n---\n\n'.join(parts)