chore: consolidate slackbot memory on letta learning-sdk (#1250)

* chore: consolidate slackbot memory on letta learning-sdk

remove home-rolled TurboPuffer-based user facts system in favor of
letta's learning-sdk which was integrated in the previous PR. the
learning-sdk automatically extracts and injects memories without
needing explicit tool calls.

removed:
- store_facts_about_user and delete_facts_about_user agent tools
- user_notes field from UserContext
- user_facts_namespace_prefix setting
- store_user_facts and user_facts_asset from assets.py
- TurboPuffer imports for user facts (kept for doc search)

note: TurboPuffer is still used for documentation search (prefect-2,
prefect-3, marvin namespaces) - only user facts storage was removed.

also adds examples/agentic_learning_demo.py showing standalone usage.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* fix: restore system prompt to agent constructor

the previous commit accidentally removed the system prompt entirely
when removing the dynamic user_notes injection. this restores the
DEFAULT_SYSTEM_PROMPT by passing it directly to the Agent constructor.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: zzstoatzz

examples/agentic_learning_demo.py

@@ -0,0 +1,59 @@
+# /// script
+# requires-python = ">=3.10"
+# dependencies = [
+#     "agentic-learning",
+#     "pydantic-ai",
+#     "pydantic-settings",
+# ]
+# [tool.uv]
+# prerelease = "allow"
+# ///
+"""
+agentic learning SDK demo
+
+this example shows how to use the agentic-learning SDK with pydantic-ai.
+the SDK automatically captures conversations and manages persistent memory.
+
+prerequisites:
+    LETTA_API_KEY and ANTHROPIC_API_KEY in .env
+
+usage:
+    uv run examples/agentic_learning_demo.py
+"""
+
+import os
+
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+
+class Settings(BaseSettings):
+    model_config = SettingsConfigDict(env_file=".env", extra="ignore")
+
+    anthropic_api_key: str
+    letta_api_key: str
+
+
+settings = Settings()
+os.environ["LETTA_API_KEY"] = settings.letta_api_key
+os.environ["ANTHROPIC_API_KEY"] = settings.anthropic_api_key
+
+from agentic_learning import learning  # noqa: E402
+from pydantic_ai import Agent  # noqa: E402
+
+agent = Agent("anthropic:claude-sonnet-4-20250514")
+
+
+def ask(message: str):
+    print(f"user: {message}\n")
+
+    with learning(agent="pydantic-ai-demo"):
+        result = agent.run_sync(message)
+        print(f"assistant: {result.output}\n")
+
+
+if __name__ == "__main__":
+    while True:
+        message = input("you: ")
+        if message.lower() in ("quit", "exit", "q"):
+            break
+        ask(message)

examples/slackbot/src/slackbot/_internal/templates.py

@@ -51,7 +51,6 @@
 - **Assume Prefect 3.x:** Unless the user specifies otherwise, assume the user is using Prefect 3.x. You can mention this assumption IF RELEVANT (e.g., "In Prefect 3.x, you would...").
 - **Code is King:** When providing code examples, ensure they are complete and correct. Use your `verify_import_statements` tool's output to guide you.
 - **Honesty Over Invention:** If your tools don't find a clear answer, say so. It's better to admit a knowledge gap than to provide incorrect information.
-- **Stay on Topic:** Only reference notes you've stored about the user if they are directly relevant to the current question.
 - **Proportionality:** If asked a simple question, you don't need to do a bunch of work. Just answer the question once you find it. However, feel free to dig into broad questions.
 
 ## CRITICAL - Removed/Deprecated Features
@@ -63,18 +62,17 @@
 If a user explicitly mentions using Prefect 2.x, that's fine, but recommend upgrading to 3.x or using workers in 2.x.
 
 ## Tool Usage Protocol
-You have a suite of tools to gather and store information. Use them methodically.
+You have a suite of tools to gather information. Use them methodically.
 
 1.  **For Technical/Conceptual Questions:** Use `research_prefect_topic`. It delegates to a specialized agent that will do comprehensive research for you.
 2.  **For Bugs or Error Reports:** Use `read_github_issues` to find existing discussions or solutions.
 3.  **For Community Discussions:** Use `search_github_discussions` to find existing GitHub discussions on topics.
-4.  **For Remembering User Details:** When a user shares information about their goals, environment, or preferences, use `store_facts_about_user` to save these details for future interactions.
-5. **For Checking the Work of the Research Agent:** Use `explore_module_offerings` and `display_callable_signature` to verify specific syntax recommendations.
-6. **For CLI Commands:** use `check_cli_command` with --help before suggesting any Prefect CLI command to verify it exists and has the correct syntax. This prevents suggesting non-existent commands.
+4. **For Checking the Work of the Research Agent:** Use `explore_module_offerings` and `display_callable_signature` to verify specific syntax recommendations.
+5. **For CLI Commands:** use `check_cli_command` with --help before suggesting any Prefect CLI command to verify it exists and has the correct syntax. This prevents suggesting non-existent commands.
    - **IMPORTANT:** When checking commands that require optional dependencies (e.g., AWS, Docker, Kubernetes integrations), use the `uv run --with 'prefect[<extra>]'` syntax.
    - Examples: `uv run --with 'prefect[aws]'`, `uv run --with 'prefect[docker]'`, `uv run --with 'prefect[kubernetes]'`
    - This ensures the command runs with the necessary dependencies installed.
-7. **For Creating GitHub Discussions (USE SPARINGLY):** Use `create_discussion_and_notify` only when:
+6. **For Creating GitHub Discussions (USE SPARINGLY):** Use `create_discussion_and_notify` only when:
    - The thread contains valuable insights, solutions, or patterns not documented elsewhere
    - You've searched both issues and discussions and found no existing coverage of the topic
    - The conversation would clearly benefit the broader Prefect community

examples/slackbot/src/slackbot/api.py

@@ -250,7 +250,6 @@ async def handle_message(payload: SlackPayload, db: Database):
 
         user_context = build_user_context(
             user_id=event.user,
-            user_question=cleaned_message,
             thread_ts=thread_ts,
             workspace_name=await get_workspace_domain(),
             channel_id=event.channel or "unknown",

examples/slackbot/src/slackbot/assets.py

@@ -4,13 +4,9 @@
 
 from prefect.assets import Asset, AssetProperties, add_asset_metadata, materialize
 from pydantic import BaseModel
-from pydantic_ai import RunContext
 from pydantic_ai.messages import ModelMessage, SystemPromptPart
-from raggy.documents import Document
-from raggy.vectorstores.tpuf import TurboPuffer
 
 from marvin import cast_async
-from slackbot.settings import settings
 from slackbot.slack import get_channel_name
 from slackbot.types import UserContext
 
@@ -74,54 +70,6 @@ def thread_summary_asset(
     )
 
 
-def user_facts_asset(user_context: UserContext) -> Asset:
-    user_id = user_context["user_id"]
-    workspace_name = user_context["workspace_name"]
-    bot_id = user_context["bot_id"]
-    return Asset(
-        key=f"slack://{workspace_name}/bot/{bot_id}/facts/{user_id}",
-        properties=AssetProperties(
-            name=f"User Facts {user_id}",
-            description=f"Facts learned about user {user_id} by bot {bot_id}",
-            owners=["slackbot"],
-        ),
-    )
-
-
-async def store_user_facts(ctx: RunContext[UserContext], facts: list[str]) -> str:
-    """Store facts extracted from a Slack thread using context for namespacing."""
-
-    with TurboPuffer(
-        namespace=f"{settings.user_facts_namespace_prefix}{ctx.deps['user_id']}"
-    ) as tpuf:
-        tpuf.upsert(documents=[Document(text=fact) for fact in facts])
-
-    user_facts = user_facts_asset(ctx.deps)
-
-    slack_thread = await slack_thread_asset(ctx.deps)
-    slackbot = slackbot_asset(ctx.deps)
-
-    @materialize(user_facts, asset_deps=[slack_thread, slackbot])
-    async def materialize_user_facts():
-        add_asset_metadata(
-            user_facts,
-            {
-                "user_id": ctx.deps["user_id"],
-                "fact_count": len(facts),
-                "timestamp": datetime.now().isoformat(),
-                "namespace": f"{settings.user_facts_namespace_prefix}{ctx.deps['user_id']}",
-                "thread_ts": ctx.deps["thread_ts"],
-                "workspace_name": ctx.deps["workspace_name"],
-                "channel_id": ctx.deps["channel_id"],
-                "bot_id": ctx.deps["bot_id"],
-                "facts": facts,
-            },
-        )
-        return f"Stored {len(facts)} facts about user {ctx.deps['user_id']} from thread {ctx.deps['thread_ts']}"
-
-    return await materialize_user_facts()
-
-
 async def summarize_thread(
     user_context: UserContext, conversation: list[ModelMessage]
 ) -> ThreadSummary:

examples/slackbot/src/slackbot/core.py

@@ -17,11 +17,8 @@
 from pydantic_ai.models.anthropic import AnthropicModel
 from pydantic_ai.providers import Provider
 from pydantic_ai.settings import ModelSettings
-from raggy.vectorstores.tpuf import TurboPuffer, query_namespace
-from turbopuffer import NotFoundError
 
 from slackbot._internal.templates import DEFAULT_SYSTEM_PROMPT
-from slackbot.assets import store_user_facts
 from slackbot.github import (
     GitHubAuthError,
     GitHubError,
@@ -136,23 +133,13 @@ def _insert():
 @task(task_run_name="build user context for {user_id}")
 def build_user_context(
     user_id: str,
-    user_question: str,
     thread_ts: str,
     workspace_name: str,
     channel_id: str,
     bot_id: str,
 ) -> UserContext:
-    try:
-        user_notes = query_namespace(
-            query_text=user_question,
-            namespace=f"{settings.user_facts_namespace_prefix}{user_id}",
-            top_k=5,
-        )
-    except NotFoundError:
-        user_notes = "<No notes found>"
     return UserContext(
         user_id=user_id,
-        user_notes=user_notes,
         thread_ts=thread_ts,
         workspace_name=workspace_name,
         channel_id=channel_id,
@@ -177,6 +164,7 @@ def create_agent(
         UserContext, str
     ](
         model=ai_model,
+        system_prompt=DEFAULT_SYSTEM_PROMPT,
         model_settings=ModelSettings(temperature=settings.temperature),
         tools=[
             research_prefect_topic,  # Tool for researching Prefect topics
@@ -189,42 +177,6 @@ def create_agent(
         deps_type=UserContext,
     )
 
-    @agent.system_prompt
-    def personality_and_maybe_notes(ctx: RunContext[UserContext]) -> str:
-        system_prompt = DEFAULT_SYSTEM_PROMPT + (
-            f"\n\nUser notes: {ctx.deps['user_notes']}"
-            if ctx.deps["user_notes"]
-            else ""
-        )
-        print(f"System prompt: {system_prompt}")
-        return system_prompt
-
-    @agent.tool
-    async def store_facts_about_user(
-        ctx: RunContext[UserContext], facts: list[str]
-    ) -> str:
-        """Store facts about the user that are useful for answering their questions."""
-        print(f"Storing {len(facts)} facts about user {ctx.deps['user_id']}")
-        # This creates an asset dependency: USER_FACTS depends on SLACK_MESSAGES
-        message = await store_user_facts(ctx, facts)
-        print(message)
-        return message
-
-    @agent.tool
-    def delete_facts_about_user(ctx: RunContext[UserContext], related_to: str) -> str:
-        """Delete facts about the user related to a specific topic."""
-        print(f"forgetting stuff about {ctx.deps['user_id']} related to {related_to}")
-        user_id = ctx.deps["user_id"]
-        with TurboPuffer(
-            namespace=f"{settings.user_facts_namespace_prefix}{user_id}"
-        ) as tpuf:
-            vector_result = tpuf.query(related_to)
-            ids = [str(v.id) for v in vector_result.rows or []]
-            tpuf.delete(ids)
-            message = f"Deleted {len(ids)} facts about user {user_id}"
-            print(message)
-            return message
-
     @agent.tool
     async def create_discussion_and_notify(
         ctx: RunContext[UserContext],

examples/slackbot/src/slackbot/prompts.py

@@ -30,7 +30,6 @@
 - **Assume Prefect 3.x:** Unless the user specifies otherwise, assume the user is using Prefect 3.x. You can mention this assumption IF RELEVANT (e.g., "In Prefect 3.x, you would...").
 - **Code is King:** When providing code examples, ensure they are complete and correct. Use your `verify_import_statements` tool's output to guide you.
 - **Honesty Over Invention:** If your tools don't find a clear answer, say so. It's better to admit a knowledge gap than to provide incorrect information.
-- **Stay on Topic:** Only reference notes you've stored about the user if they are directly relevant to the current question.
 - **Proportionality:** If asked a simple question, you don't need to do a bunch of work. Just answer the question once you find it. However, feel free to dig into broad questions.
 
 ## CRITICAL - Removed/Deprecated Features
@@ -42,18 +41,17 @@
 If a user explicitly mentions using Prefect 2.x, that's fine, but recommend upgrading to 3.x or using workers in 2.x.
 
 ## Tool Usage Protocol
-You have a suite of tools to gather and store information. Use them methodically.
+You have a suite of tools to gather information. Use them methodically.
 
 1.  **For Technical/Conceptual Questions:** Use `research_prefect_topic`. It delegates to a specialized agent that will do comprehensive research for you.
 2.  **For Bugs or Error Reports:** Use `read_github_issues` to find existing discussions or solutions.
 3.  **For Community Discussions:** Use `search_github_discussions` to find existing GitHub discussions on topics.
-4.  **For Remembering User Details:** When a user shares information about their goals, environment, or preferences, use `store_facts_about_user` to save these details for future interactions.
-5. **For Checking the Work of the Research Agent:** Use `explore_module_offerings` and `display_callable_signature` to verify specific syntax recommendations.
-6. **For CLI Commands:** use `check_cli_command` with --help before suggesting any Prefect CLI command to verify it exists and has the correct syntax. This prevents suggesting non-existent commands.
+4. **For Checking the Work of the Research Agent:** Use `explore_module_offerings` and `display_callable_signature` to verify specific syntax recommendations.
+5. **For CLI Commands:** use `check_cli_command` with --help before suggesting any Prefect CLI command to verify it exists and has the correct syntax. This prevents suggesting non-existent commands.
    - **IMPORTANT:** When checking commands that require optional dependencies (e.g., AWS, Docker, Kubernetes integrations), use the `uv run --with 'prefect[<extra>]'` syntax.
    - Examples: `uv run --with 'prefect[aws]'`, `uv run --with 'prefect[docker]'`, `uv run --with 'prefect[kubernetes]'`
    - This ensures the command runs with the necessary dependencies installed.
-7. **For Creating GitHub Discussions (USE SPARINGLY):** Use `create_discussion_and_notify` only when:
+6. **For Creating GitHub Discussions (USE SPARINGLY):** Use `create_discussion_and_notify` only when:
    - The thread contains valuable insights, solutions, or patterns not documented elsewhere
    - You've searched both issues and discussions and found no existing coverage of the topic
    - The conversation would clearly benefit the broader Prefect community

examples/slackbot/src/slackbot/settings.py

@@ -1,6 +1,6 @@
 import os
 from pathlib import Path
-from typing import ClassVar, Literal
+from typing import ClassVar
 
 from prefect.blocks.system import Secret
 from prefect.exceptions import ObjectNotFound
@@ -62,14 +62,6 @@ def validate_log_level(cls, v: str) -> str:
         description="Name of the Prefect secret block containing Letta API key",
     )
 
-    vector_store_type: Literal["turbopuffer"] = Field(
-        default="turbopuffer", description="Type of vector store to use"
-    )
-    user_facts_namespace_prefix: str = Field(
-        default="user-facts-",
-        description="Prefix for user facts namespaces in vector store",
-    )
-
     # Development settings
     test_mode: bool = Field(
         default=False, description="Enable test mode with auto-reload"

examples/slackbot/src/slackbot/types.py

@@ -3,7 +3,6 @@
 
 class UserContext(TypedDict):
     user_id: str
-    user_notes: str
     thread_ts: str
     workspace_name: str
     channel_id: str