fix(mem0-plugin): store assistant-authored summaries with role="assistant" (#6316)

.claude-plugin/marketplace.json

@@ -12,7 +12,7 @@
       "name": "mem0",
       "source": "./integrations/mem0-plugin",
       "description": "Mem0 memory layer for AI applications. Add persistent memory, personalization, and semantic search to Claude workflows.",
-      "version": "0.2.12"
+      "version": "0.2.13"
     }
   ]
 }

.cursor-plugin/marketplace.json

@@ -12,7 +12,7 @@
       "name": "mem0",
       "source": "./integrations/mem0-plugin",
       "description": "Mem0 memory layer for AI applications. Add persistent memory, personalization, and semantic search.",
-      "version": "0.2.12"
+      "version": "0.2.13"
     }
   ]
 }

docs/changelog/sdk.mdx

@@ -1825,6 +1825,15 @@ A full-featured command-line interface for Mem0, available in both Python and No
 <Tabs>
 <Tab title="Mem0 Plugin">
 
+<Update label="2026-07-14" description="mem0-plugin v0.2.13">
+
+**Fixes:**
+- **Assistant messages no longer stored as your own:** The session-summary hook (fires at the end of every assistant turn) and the post-compaction hook were sending the assistant's own message to Mem0 tagged `role: "user"`. Because Mem0 extracts *facts about the user* from each message and uses `role` to decide who spoke, the assistant's first-person prose was being saved as the human's stated preferences — "I recommend we drop Redis" became `User prefers dropping Redis entirely`. Both hooks now send `role: "assistant"`, so the same session is stored as `Assistant recommended...`. Affects Claude Code, Cursor, Codex, and Antigravity, which share these hooks.
+
+Existing memories written by the previous versions are not rewritten. If your memories contain preferences you never expressed, delete them — the plugin will not recreate them.
+
+</Update>
+
 <Update label="2026-06-30" description="mem0-plugin v0.2.12">
 
 **New Features:**
@@ -2148,6 +2157,15 @@ Initial release of the Mem0 plugin for Claude Code and Cursor, followed by Codex
 
 <Tab title="Antigravity">
 
+<Update label="2026-07-14" description="Antigravity plugin v0.1.5">
+
+**Fixes:**
+- **Assistant messages no longer stored as your own:** The session-summary hook (fires at the end of every assistant turn) and the post-compaction hook were sending the assistant's own message to Mem0 tagged `role: "user"`. Because Mem0 extracts *facts about the user* from each message and uses `role` to decide who spoke, the assistant's first-person prose was being saved as the human's stated preferences — "I recommend we drop Redis" became `User prefers dropping Redis entirely`. Both hooks now send `role: "assistant"`, so the same session is stored as `Assistant recommended...`.
+
+Existing memories written by the previous versions are not rewritten. If your memories contain preferences you never expressed, delete them — the plugin will not recreate them.
+
+</Update>
+
 <Update label="2026-06-30" description="Antigravity plugin v0.1.4">
 
 **New Features:**

docs/integrations/antigravity.mdx

@@ -65,7 +65,9 @@ The plugin uses the same shell scripts as Claude Code, Cursor, and Codex: hooks
 | **User prompt** | `UserPromptSubmit` | Searches relevant memories before each message |
 | **Pre-tool** | `PreToolUse` | Blocks MEMORY.md writes, enforces `user_id`/`app_id` on mem0 tools |
 | **Post-tool** | `PostToolUse` | Tracks stats, scans bash errors for related memories |
-| **Stop** | `Stop` | Stores a session summary when the session ends |
+| **Stop** | `Stop` | Stores a session summary at the end of every assistant turn (not just at session end) |
+
+What you type is stored as yours. What the agent produces — session summaries and compaction summaries — is stored as the assistant's, so its suggestions never become your stated preferences.
 
 ## Troubleshooting
 

docs/integrations/claude-code.mdx

@@ -143,9 +143,11 @@ When installed via the plugin marketplace, Mem0 hooks into Claude Code's lifecyc
 | **User prompt** | `UserPromptSubmit` | Searches relevant memories before each message; skips short prompts |
 | **Pre-tool (3 handlers)** | `PreToolUse` | Blocks MEMORY.md writes; enforces `user_id`/`app_id` on mem0 tool calls; scans files being read for relevant memory context |
 | **Post-tool** | `PostToolUse` | Tracks stats, scans bash errors for related memories |
-| **Stop** | `Stop` | Stores a session summary when the session ends |
+| **Stop** | `Stop` | Stores a session summary at the end of every assistant turn (not just at session end) |
 | **Pre-compact** | `PreCompact` | Stores a summary before the context is compacted |
 
+What you type is stored as yours. What Claude produces — session summaries and compaction summaries — is stored as the assistant's, so its suggestions never become your stated preferences.
+
 ## Example Workflow
 
 ```text
@@ -153,16 +155,17 @@ When installed via the plugin marketplace, Mem0 hooks into Claude Code's lifecyc
 You: Let's refactor the auth module to use JWT tokens instead of sessions.
 
 # Claude searches memories, finds nothing relevant, proceeds with the work.
-# After completing the task, Mem0 stores:
+# Mem0 stores what you said as yours:
+#   - Your preference: "Prefers TypeScript, uses ESLint"
+# ...and what Claude did as the assistant's, in the session summary:
 #   - Decision: "Migrated auth from sessions to JWT tokens"
 #   - Files modified: auth/middleware.ts, auth/token.ts
-#   - User preference: "Prefers TypeScript, uses ESLint"
 
 # Session 2 (days later): Related work
 You: Add refresh token rotation to the auth system.
 
 # Claude searches memories, retrieves the JWT migration context.
-# Knows the file structure, decisions made, and user preferences.
+# Knows the file structure, decisions made, and your stated preferences.
 # Continues seamlessly without re-explaining the codebase.
 ```
 

docs/integrations/codex.mdx

@@ -125,20 +125,23 @@ When installed via the plugin marketplace, Mem0 hooks into Codex's lifecycle to
 | **User prompt** | `UserPromptSubmit` | Searches relevant memories before each message |
 | **Pre-tool (3 handlers)** | `PreToolUse` | Blocks MEMORY.md writes; enforces `user_id`/`app_id` on mem0 tool calls; scans files being read for relevant memory context |
 | **Post-tool** | `PostToolUse` | Tracks stats, scans bash errors for related memories |
-| **Stop** | `Stop` | Stores a session summary when the session ends |
+| **Stop** | `Stop` | Stores a session summary at the end of every assistant turn (not just at session end) |
 | **Pre-compact** | `PreCompact` | Stores a summary before the context is compacted |
 
+What you type is stored as yours. What Codex produces — session summaries and compaction summaries — is stored as the assistant's, so its suggestions never become your stated preferences.
+
 ## Example Workflow
 
 ```text
 # Task 1: Setting up a new service
 You: Create a REST API for the notifications service using Express and TypeScript.
 
-# Codex searches memories, finds user preferences from prior tasks.
+# Codex searches memories, finds your preferences from prior tasks.
-# After completing the task, Mem0 stores:
+# Mem0 stores what you said as yours:
+#   - Your preference: "Prefers explicit error types over generic catch-all"
+# ...and what Codex did as the assistant's, in the session summary:
 #   - Decision: "Notifications service uses Express + TypeScript + Zod validation"
 #   - Convention: "All API routes follow /api/v1/{resource} pattern"
-#   - Preference: "User prefers explicit error types over generic catch-all"
 
 # Task 2 (days later): Extending the service
 You: Add WebSocket support for real-time notification delivery.

docs/integrations/cursor.mdx

@@ -96,10 +96,11 @@ Once installed, the following tools are available in every Cursor session:
 You: The API endpoint /users is taking 3 seconds. Help me optimize it.
 
 # Cursor agent searches memories, proceeds with investigation.
-# After completing the task, Mem0 stores:
+# Mem0 stores what you said as yours:
+#   - Your preference: "Prefers query-level fixes over caching"
+# ...and what the agent did as the assistant's:
 #   - Learning: "N+1 query in UserService.getAll(): fixed with eager loading"
 #   - Decision: "Added database index on users.email column"
-#   - Preference: "User prefers query-level fixes over caching"
 
 # Session 2 (next week): Similar issue
 You: The /orders endpoint is also slow, same pattern as before.

integrations/mem0-plugin/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "mem0",
-  "version": "0.2.12",
+  "version": "0.2.13",
   "description": "Persistent memory for Claude Code. Remembers decisions, patterns, and preferences across sessions.",
   "author": {
     "name": "Mem0",

integrations/mem0-plugin/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "mem0",
-  "version": "0.2.12",
+  "version": "0.2.13",
   "description": "Persistent memory for Codex. Remembers decisions, patterns, and preferences across sessions.",
   "author": {
     "name": "Mem0",

integrations/mem0-plugin/.cursor-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "mem0",
-  "version": "0.2.12",
+  "version": "0.2.13",
   "description": "Mem0 memory layer for AI applications. Add persistent memory, personalization, and semantic search using the Mem0 Platform MCP server.",
   "author": {
     "name": "Mem0",

integrations/mem0-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "id": "mem0",
   "name": "mem0",
-  "version": "0.1.4",
+  "version": "0.1.5",
   "description": "Persistent semantic memory for Antigravity agents. Cross-session, user-level recall via the Mem0 Platform MCP server. 16 slash commands, lifecycle hooks for auto-capture and metadata enforcement.",
   "author": { "name": "Mem0", "email": "support@mem0.ai" },
   "publisher": "mem0ai",

integrations/mem0-plugin/scripts/capture_compact_summary.py

@@ -104,8 +104,11 @@ def store_summary(api_key: str, summary: str, user_id: str, session_id: str, pro
     }
     if branch:
         metadata["branch"] = branch
+    # The compact summary is model-authored prose, in the first person and with no
+    # framing to mark it as such. Under role="user" mem0 reads "I recommend X" as
+    # the human saying it and stores "User recommends X".
     body = {
-        "messages": [{"role": "user", "content": summary}],
+        "messages": [{"role": "assistant", "content": summary}],
         "user_id": user_id,
         "app_id": project_id,
         "metadata": metadata,

integrations/mem0-plugin/scripts/capture_session_summary.py

@@ -175,8 +175,12 @@ def store_summary(
     if files:
         metadata["files_touched"] = files[:20]
 
+    # summary_prompt wraps the assistant's own last message. Mem0 extracts "facts
+    # about the user" from each message and role is the only signal telling it who
+    # spoke, so role="user" here turns Claude's opinions into the human's stated
+    # preferences ("User prefers dropping Redis...").
     body = {
-        "messages": [{"role": "user", "content": summary_prompt}],
+        "messages": [{"role": "assistant", "content": summary_prompt}],
         "user_id": user_id,
         "app_id": project_id,
         "run_id": session_id,

integrations/mem0-plugin/tests/test_message_roles.py

@@ -0,0 +1,144 @@
+"""Regression tests: assistant-authored text must never be posted as role="user".
+
+The Stop hook (capture_session_summary) and the post-compact hook
+(capture_compact_summary) both ship *model-authored* prose to
+POST /v3/memories/add/. Mem0's fact extractor renders each message as
+"{role}: {content}" and is instructed to extract "facts and preferences about
+the user" — so role is the only signal separating what the human said from what
+Claude said.
+
+Posting Claude's own words under role="user" made the extractor read Claude's
+first-person prose ("I recommend pgvector", "I found the bug in auth.py") as the
+*human's* statements and store them under their user_id. The Stop hook fires on
+every assistant turn, so this corrupted memory on nearly every message.
+"""
+
+from __future__ import annotations
+
+import json
+
+
+class _FakeResp:
+    status = 200
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *_):
+        return False
+
+
+def _capture(monkeypatch, module):
+    """Patch urlopen so store_summary posts nowhere; capture the request body."""
+    captured: dict = {}
+
+    def fake_urlopen(req, timeout=0):
+        captured["body"] = json.loads(req.data.decode("utf-8"))
+        return _FakeResp()
+
+    monkeypatch.setattr(module.urllib.request, "urlopen", fake_urlopen)
+    return captured
+
+
+# Claude's own voice — first-person prose that must never be attributed to the human.
+ASSISTANT_PROSE = (
+    "I traced the root cause to auth.py and I recommend we switch to pgvector "
+    "for the vector store. I'll refactor the session handler next."
+)
+
+
+def test_session_summary_posts_assistant_prose_as_assistant(monkeypatch):
+    """Stop hook: the last assistant message must be tagged role="assistant"."""
+    import capture_session_summary as css
+
+    captured = _capture(monkeypatch, css)
+
+    css.store_summary(
+        api_key="test-key",
+        summary_prompt=css.build_summary_prompt(ASSISTANT_PROSE, []),
+        user_id="u1",
+        session_id="s1",
+        project_id="p1",
+        branch="main",
+        files=[],
+    )
+
+    messages = captured["body"]["messages"]
+    for msg in messages:
+        if ASSISTANT_PROSE in msg["content"]:
+            assert msg["role"] == "assistant", (
+                "Claude's own words were posted as role='user' — mem0 will extract "
+                "them as facts about the human. Got role=%r" % msg["role"]
+            )
+            break
+    else:
+        raise AssertionError("assistant prose never made it into the payload")
+
+
+def test_compact_summary_posts_assistant_prose_as_assistant(monkeypatch):
+    """Post-compact hook: the compact summary is model-authored, not user-authored."""
+    import capture_compact_summary as ccs
+
+    captured = _capture(monkeypatch, ccs)
+
+    ccs.store_summary(
+        api_key="test-key",
+        summary=ASSISTANT_PROSE,
+        user_id="u1",
+        session_id="s1",
+        project_id="p1",
+        branch="main",
+    )
+
+    messages = captured["body"]["messages"]
+    for msg in messages:
+        if ASSISTANT_PROSE in msg["content"]:
+            assert msg["role"] == "assistant", (
+                "Compact summary (written by Claude) was posted as role='user'. Got role=%r" % msg["role"]
+            )
+            break
+    else:
+        raise AssertionError("assistant prose never made it into the payload")
+
+
+def test_no_user_role_message_carries_assistant_prose(monkeypatch):
+    """Belt and braces: no user-role message may contain the assistant's words."""
+    import capture_session_summary as css
+
+    captured = _capture(monkeypatch, css)
+
+    css.store_summary(
+        api_key="test-key",
+        summary_prompt=css.build_summary_prompt(ASSISTANT_PROSE, ["auth.py"]),
+        user_id="u1",
+        session_id="s1",
+        project_id="p1",
+        branch="main",
+        files=["auth.py"],
+    )
+
+    for msg in captured["body"]["messages"]:
+        if msg["role"] == "user":
+            assert ASSISTANT_PROSE not in msg["content"], (
+                "A user-role message carries Claude's prose — this is the misattribution bug."
+            )
+
+
+def test_auto_capture_preserves_real_roles():
+    """auto_capture is the reference: it must pass roles through untouched."""
+    import auto_capture
+
+    lines = [
+        json.dumps({"type": "user", "message": {"role": "user", "content": "why is the build failing on main?"}}),
+        json.dumps(
+            {
+                "type": "assistant",
+                "message": {"role": "assistant", "content": [{"type": "text", "text": ASSISTANT_PROSE}]},
+            }
+        ),
+    ]
+
+    messages = auto_capture.extract_recent_exchanges(lines)
+
+    assert [m["role"] for m in messages] == ["user", "assistant"]
+    assert ASSISTANT_PROSE in messages[1]["content"]