AG2 integration

AG2 agents using Zep maintain context across conversations and access a temporal knowledge graph. The zep-ag2 package provides search and data tools. It also provides an automatic system-message path for trusted content.

Use the Zep search tools for context that can contain end-user or third-party data. Use the automatic loop’s system-message injection only for fully trusted, application-authored context.

Core benefits

• Automatic memory loop for trusted content: attach_to_agent persists every message and can refresh the system message with application-authored, trusted context
• Persistent memory: Conversations and extracted knowledge persist across sessions
• System message injection for trusted content: Relevant context can be added to an agent’s system message before it responds
• Knowledge graph access: Search and write to Zep’s temporal knowledge graph from AG2 agents
• Tool-based access: Register Zep search and add operations as AG2 tools the agent invokes on demand

How it works

AG2 has no native memory interface, so the integration provides three ways to give an agent memory:

• Automatic memory loop — ZepMemoryManager.attach_to_agent(agent) registers hooks on ConversableAgent that persist every message the agent receives and sends, and refresh its system message with relevant context on each turn. Use this path only for fully trusted, application-authored context.
• System message injection for trusted content — ZepMemoryManager and ZepGraphMemoryManager can add trusted context to an agent’s system message.
• Tools — factory functions return AG2-compatible tools the model can call mid-conversation to search memory or write new data. Tools execute synchronously (AG2’s execution model) while bridging to the async Zep SDK internally, so you pass an AsyncZep client.

Use tools for end-user or third-party context. Attach the automatic loop only when all stored content is application-authored and trusted.

Installation

$
pip install zep-ag2

Set up your environment variables:

$
export ZEP_API_KEY="your-zep-api-key"
$
export OPENAI_API_KEY="your-openai-api-key"

Automatic memory loop for trusted deployments

attach_to_agent writes conversation content to memory and later inserts that memory into a system message. Do not use this path when a user or external source can influence stored content.

1
import os
2
from autogen import AssistantAgent, UserProxyAgent, LLMConfig
3
from zep_cloud.client import AsyncZep
4
from zep_ag2 import ZepMemoryManager
5
6
zep = AsyncZep(api_key=os.environ["ZEP_API_KEY"])
7
8
llm_config = LLMConfig(
9
    {"model": "gpt-5.6-terra", "api_key": os.environ["OPENAI_API_KEY"]}
10
)
11
12
assistant = AssistantAgent(
13
    name="assistant",
14
    llm_config=llm_config,
15
    system_message="You are a helpful assistant with long-term memory.",
16
)
17
user_proxy = UserProxyAgent(
18
    name="user",
19
    human_input_mode="NEVER",
20
    code_execution_config=False,
21
)
22
23
# Trusted-only path: attach_to_agent inserts stored content into a system message.
24
manager = ZepMemoryManager(zep, user_id="user123", session_id="session456")
25
manager.attach_to_agent(assistant)
26
27
# Every message the assistant receives is persisted and used to refresh its
28
# system message; every reply it sends is persisted automatically too.
29
user_proxy.initiate_chat(assistant, message="My name is Alice.")

attach_to_agent(agent) registers two hooks on AG2’s ConversableAgent:

• process_last_received_message fires for every message the agent receives. It persists the message and retrieves fresh context (via process_user_message internally), then replaces the agent’s system message with its original text plus the rendered context template. The hook returns the message content unmodified — it is a side channel, not a message transform.
• process_message_before_send fires for every message the agent sends. It persists the outgoing message as an assistant message and returns it unchanged.

Both hooks wrap their entire body in error handling, so a Zep outage never breaks the agent’s conversation loop — on failure, the incoming hook skips the system-message update and the outgoing hook skips persistence, in both cases still returning the message unchanged.

attach_to_agent is optional and additive: enrich_system_message and add_messages remain available for manual control, for example to persist only some turns or inject context at a different point than “on receive”.

Manual system-message injection for trusted deployments

Use ZepMemoryManager.enrich_system_message only when all stored content is fully trusted and application-authored:

1
import asyncio
2
import os
3
from autogen import AssistantAgent, UserProxyAgent, LLMConfig
4
from zep_cloud.client import AsyncZep
5
from zep_ag2 import ZepMemoryManager, ensure_user, ensure_thread, register_all_tools
6
7
async def main():
8
    zep = AsyncZep(api_key=os.environ["ZEP_API_KEY"])
9
    user_id = "user123"
10
    session_id = "session456"
11
12
    # Optional: provision out-of-band so failures surface before the first turn
13
    await ensure_user(zep, user_id=user_id, first_name="Jane")
14
    await ensure_thread(zep, thread_id=session_id, user_id=user_id)
15
16
    llm_config = LLMConfig(
17
        {"model": "gpt-5.6-terra", "api_key": os.environ["OPENAI_API_KEY"]}
18
    )
19
20
    assistant = AssistantAgent(
21
        name="assistant",
22
        llm_config=llm_config,
23
        system_message="You are a helpful assistant with long-term memory.",
24
    )
25
    user_proxy = UserProxyAgent(
26
        name="user",
27
        human_input_mode="NEVER",
28
        code_execution_config=False,
29
        is_termination_msg=lambda msg: "TERMINATE" in (msg.get("content") or ""),
30
    )
31
32
    # Trusted-only path: never pass end-user or third-party memory here.
33
    memory_mgr = ZepMemoryManager(zep, user_id=user_id, session_id=session_id)
34
    await memory_mgr.enrich_system_message(assistant, query="conversation topic")
35
36
    # Register Zep memory tools — AG2 calls them automatically
37
    register_all_tools(assistant, user_proxy, zep, user_id=user_id, session_id=session_id)
38
39
    user_proxy.initiate_chat(assistant, message="What do you remember about me?")
40
41
asyncio.run(main())

ZepMemoryManager also exposes process_user_message() to persist a user turn and retrieve context in one call, get_memory_context() to retrieve the formatted context string directly, add_messages() to persist conversation turns, and get_session_facts() to read the thread’s context block.

Provisioning

The manager creates the Zep user and (when a session_id is set) thread lazily on the first memory-path call — process_user_message, get_memory_context, enrich_system_message, add_messages, or the attach_to_agent hooks. Creation is idempotent and cached per manager instance, so no pre-creation step is required.

Pass first_name, last_name, and email so Zep can anchor the user’s identity node in the graph, and on_created to run one-time setup (ontology, custom instructions) only when the user is newly created:

1
async def setup_new_user(zep, user_id: str) -> None:
2
    ...  # one-time setup: ontology, custom instructions
3
4
manager = ZepMemoryManager(
5
    zep,
6
    user_id="user123",
7
    session_id="session456",
8
    first_name="Jane",
9
    last_name="Smith",
10
    email="[email protected]",
11
    on_created=setup_new_user,
12
)

The lazy path never raises into a memory-path method: a genuine provisioning failure (or an on_created hook failure) is logged and swallowed. To surface provisioning failures loudly — for example during account onboarding, before the first turn — call ensure_user and ensure_thread out-of-band:

1
from zep_ag2 import ensure_user, ensure_thread
2
3
await ensure_user(zep, user_id="user123", first_name="Jane", on_created=setup_new_user)
4
await ensure_thread(zep, thread_id="session456", user_id="user123")

Both helpers are idempotent and return True only when the resource is newly created.

A ZepMemoryManager is scoped to one (user_id, session_id) pair for the lifetime of the instance — create one manager per user/thread rather than sharing an instance across users.

Tool integration

Register Zep operations as AG2 tools so the agent can search memory or write new data during a conversation. register_all_tools wires up the full set in one call, or use the individual factories for finer control:

1
from zep_ag2 import create_search_graph_tool, create_add_graph_data_tool
2
3
# Create tools bound to a user's knowledge graph
4
search_tool = create_search_graph_tool(zep, user_id="user123")
5
add_tool = create_add_graph_data_tool(zep, user_id="user123")
6
7
# Register with AG2's decorator pattern
8
assistant.register_for_llm(description="Search knowledge graph")(search_tool)
9
user_proxy.register_for_execution()(search_tool)
10
11
assistant.register_for_llm(description="Add to knowledge graph")(add_tool)
12
user_proxy.register_for_execution()(add_tool)

Available tool factories:

• create_search_memory_tool(client, user_id, session_id=None, *, pinned_params=None, hidden_params=None, search_filters=None, bfs_origin_node_uuids=None, scope=None, limit=None) — searches the user’s graph
• create_add_memory_tool(client, user_id, session_id=None) — routes to the thread when a session_id is set, otherwise writes to the user’s graph
• create_search_graph_tool(client, user_id=None, graph_id=None, *, pinned_params=None, hidden_params=None, search_filters=None, bfs_origin_node_uuids=None, scope=None, limit=None) — search the knowledge graph
• create_add_graph_data_tool(client, user_id=None, graph_id=None) — add data to the knowledge graph
• register_all_tools(agent, executor, client, user_id, ...) — register all tools at once

Graph tools are bound to either a user_id (the user’s personal graph) or a graph_id (a shared standalone graph), not both.

Search tool parameters

The search tool factories follow a pin-or-expose pattern: every graph.search parameter is exposed to the model by default, each with a typed schema and documented default. Letting the model choose the scope and reranker per query produces better retrieval than a single fixed configuration; pin parameters when you need deterministic behavior instead.

Use pinned_params to fix a parameter to a constant (hidden from the model), or hidden_params to remove it from the schema without pinning (Zep’s server-side default applies):

1
# Model chooses scope/reranker/limit/mmr_lambda/center_node_uuid freely (default)
2
tool = create_search_graph_tool(zep, user_id="user123")
3
4
# Pin scope to "nodes" and limit to 5 — hidden from the model, always sent as given
5
tool = create_search_graph_tool(
6
    zep, user_id="user123", pinned_params={"scope": "nodes", "limit": 5}
7
)
8
9
# Hide reranker entirely — omitted from the schema and the SDK call
10
tool = create_search_graph_tool(zep, user_id="user123", hidden_params={"reranker"})

The legacy scope and limit keyword arguments pin (and hide) the corresponding parameter — equivalent to passing them via pinned_params. search_filters and bfs_origin_node_uuids are constructor-only and never exposed to the model.

Knowledge graph memory

Use ZepGraphMemoryManager to work with a shared knowledge graph that multiple agents can read and write:

1
from zep_ag2 import ZepGraphMemoryManager
2
3
graph_mgr = ZepGraphMemoryManager(zep, graph_id="company_knowledge")
4
5
# Add data to the graph
6
await graph_mgr.add_data("Project Alpha uses Python and React.", data_type="text")
7
8
# Search the graph
9
results = await graph_mgr.search("What technologies does Project Alpha use?", limit=5, scope="edges")

ZepGraphMemoryManager.search() accepts scope values edges, nodes, and episodes and returns structured result dicts for programmatic use. This is distinct from the search tool schema above, which exposes six scopes to the model and returns formatted strings.

Use AG2 search tools or place these results through your provider’s data channel for end-user or third-party graph content. enrich_system_message is only for a fully trusted, application-authored graph.

Custom context retrieval

By default, context is retrieved via thread.get_user_context(...) (or, inside process_user_message, via thread.add_messages(..., return_context=True)). Pass context_builder to replace this with custom logic — for example a filtered graph search, or a different graph entirely:

1
from zep_ag2.memory import ContextInput
2
3
async def my_builder(ctx: ContextInput) -> str | None:
4
    results = await ctx.zep.graph.search(
5
        user_id=ctx.user_id,
6
        query=ctx.user_message,
7
        scope="edges",
8
    )
9
    if not results.edges:
10
        return None
11
    return "\n".join(edge.fact for edge in results.edges)
12
13
manager = ZepMemoryManager(
14
    zep, user_id="user123", session_id="session456", context_builder=my_builder,
15
)

The builder receives a single frozen ContextInput:

If the builder raises, a warning is logged and context injection is skipped for that call — the builder never raises into process_user_message, get_memory_context, or enrich_system_message. Inside process_user_message, persistence and the builder run concurrently with per-side isolation: a builder failure never blocks the message from being persisted, and a persistence failure never prevents the builder’s context from being returned.

Customizing the injected context template

Retrieved context (from the default retrieval or a context_builder) is wrapped in context_template before injection into the agent’s system message. The default DEFAULT_CONTEXT_TEMPLATE wraps the context in <ZEP_CONTEXT> tags with a short preamble. Override it with your own wording, as long as it contains a literal {context} placeholder:

1
manager = ZepMemoryManager(
2
    zep, user_id="user123", session_id="session456",
3
    context_template="Relevant background:\n{context}",
4
)

The template uses plain string replacement (template.replace("{context}", ...)), never str.format. This prevents format-string interpretation of {, }, or %. It does not prevent the model from following instructions in the retrieved content.

Query memory

You can read memory directly, outside of agent tool calls:

1
# Formatted context block for the user/thread (optionally biased by a query)
2
context = await memory_mgr.get_memory_context(query="project status", limit=5)
3
4
# Facts extracted from the current session
5
facts = await memory_mgr.get_session_facts()
6
7
# Structured search over a knowledge graph
8
results = await graph_mgr.search("Project Alpha", limit=5, scope="edges")

Search result structure

The tool factories return human-readable strings formatted for the model, with formatting that adapts to the search scope. ZepGraphMemoryManager.search() returns a list of structured result dicts for programmatic use; the fields depend on the scope:

Memory vs tools

The integration supports three patterns:

Use tools as the default for context that users or external sources can influence.

Configuration options

ZepMemoryManager

• ZepMemoryManager(client, user_id, session_id=None, *, first_name=None, last_name=None, email=None, on_created=None, context_builder=None, context_template=DEFAULT_CONTEXT_TEMPLATE) — initialize with a Zep client and user identity; the configuration arguments are keyword-only
• attach_to_agent(agent) — register the automatic inject and persist loop
• process_user_message(user_message, *, agent=None) — persist a user turn and retrieve context in one call
• ensure_user_and_thread() — lazily provision the user and thread; returns False on failure, never raises
• enrich_system_message(agent, query=None, limit=5) — inject memory context into an agent
• get_memory_context(query=None, limit=5) — return the formatted context string
• add_messages(messages) — store messages in the Zep thread
• get_session_facts() — read the thread’s context block

ZepGraphMemoryManager

• ZepGraphMemoryManager(client, graph_id) — initialize with a graph ID
• search(query, limit=5, scope="edges") — search the graph (scope: edges, nodes, episodes)
• add_data(data, data_type="text") — add data to the graph (data_type: text, json, message)
• enrich_system_message(agent, query=None, limit=5) — inject graph context into an agent

Provisioning helpers

• ensure_user(client, *, user_id, first_name=None, last_name=None, email=None, on_created=None) — idempotently create a Zep user; returns True only when newly created
• ensure_thread(client, *, thread_id, user_id) — idempotently create a Zep thread; returns True only when newly created

Size limits

Zep rejects over-long direct SDK payloads with an HTTP 400. The AG2 integration truncates before calling Zep, logging only the before and after lengths (never the content):

• Thread messages: truncated to 4,000 characters, a safety margin under Zep’s 4,096-character thread-message limit
• Graph data (add_data, create_add_graph_data_tool): truncated to 9,900 characters, a safety margin under Zep’s 10,000-character graph.add limit

Best practices

• Pass an AsyncZep client — tools bridge to it on a shared background event loop, so reuse a single instance
• Attach one agent per Zep thread — attaching two managers with the same session_id double-persists every turn with conflicting roles
• Bind tools to one target — a user_id for personal memory or a graph_id for shared knowledge, never both
• Use tools for untrusted context. Attach the automatic loop only when all stored content is application-authored and trusted.
• Allow time for indexing — Zep extracts knowledge asynchronously, so data added during a turn is not instantly searchable

Next steps

• Explore customizing graph structure for advanced knowledge organization
• Learn about searching the graph and how to tune search
• See code examples for additional patterns