Mastra integration

Mastra agents using Zep gain long-term memory backed by a temporal knowledge graph. The @getzep/zep-mastra package provides two complementary surfaces:

• Automatic memory for trusted content — createZepProcessors builds a ZepInputProcessor/ZepOutputProcessor pair that plugs into Mastra’s native inputProcessors/outputProcessors pipeline. The input processor inserts context into a system message.
• Tools — createZepToolset builds zepRemember/zepSearch/zepContext tools that let the model decide when to persist or recall. Use retrieval tools for end-user or third-party context.

Core benefits

• Automatic memory loop for trusted content: Processors can inject application-authored context and persist each completed turn
• Model-in-the-loop tools: zepRemember, zepSearch, and zepContext drop straight into an Agent’s tools record
• Per-call identity: Resolve userId/threadId from Mastra’s requestContext so one processor or tool instance serves many end users
• User and standalone graphs: Bind to a user’s personal graph or a shared knowledge base
• Graceful degradation: A Zep outage is logged and surfaced as a non-fatal result — it never crashes the host agent

How it works

The processors sit on opposite sides of the model call:

1. ZepInputProcessor runs before the model is called. It extracts the latest user message, retrieves a Zep context block (thread.getUserContext, or a custom contextBuilder), wraps it with contextTemplate/formatContext, and injects it as a system message.
2. ZepOutputProcessor runs after the model responds. It persists the completed turn — the latest user message plus the assistant’s response — to the bound thread via a single thread.addMessages call. The assistant text persisted is the final step’s text; when generation ends mid-tool-loop (finishReason === "tool-calls"), the user message is still persisted.

Because injection and persistence sit on opposite sides of the model call, it’s safe to enable both processors together — they don’t interfere with each other. Every Zep call is wrapped: a missing threadId or any Zep failure degrades gracefully — messages pass through unchanged, a warning is logged — and the input processor never calls abort() or throws into the agent loop.

Zep is a temporal knowledge graph, not a row-oriented message store, so the package exposes Zep’s two real operations — persist and retrieve — through processors and tools rather than a MastraStorage adapter, which would require CRUD operations a temporal knowledge graph can’t honor faithfully.

Installation

$
npm install @getzep/zep-mastra @getzep/zep-cloud @mastra/core

Set up your environment variables:

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

Automatic memory for trusted deployments

createZepProcessors inserts retrieved context into a system message. Use this processor pair only when all stored content is fully trusted and application-authored:

1
import { ZepClient } from "@getzep/zep-cloud";
2
import { Agent } from "@mastra/core/agent";
3
import { createZepProcessors, ensureZepUserAndThread } from "@getzep/zep-mastra";
4
5
const client = new ZepClient({ apiKey: process.env.ZEP_API_KEY! });
6
const userId = "user-123";
7
const threadId = "thread-abc";
8
9
// 1. Provision the Zep user + thread before the first turn.
10
await ensureZepUserAndThread({ client, userId, threadId, firstName: "Jane", lastName: "Smith" });
11
12
// 2. Trusted-only path: the input processor inserts context into a system message.
13
const { inputProcessor, outputProcessor } = createZepProcessors({ client, userId, threadId });
14
15
// 3. Attach to a Mastra agent (id and name are both required by Mastra).
16
const agent = new Agent({
17
  id: "memory-agent",
18
  name: "Memory Agent",
19
  instructions: "You have long-term memory about the user. Use it to personalize replies.",
20
  model: "openai/gpt-5.6-terra",
21
  inputProcessors: [inputProcessor],
22
  outputProcessors: [outputProcessor],
23
});

Customizing context injection

By default the input processor retrieves the context block with thread.getUserContext and wraps it in DEFAULT_CONTEXT_TEMPLATE — the canonical <ZEP_CONTEXT> wrapper shared by Zep’s framework integrations. Three options change that, in increasing order of control:

1
const { inputProcessor, outputProcessor } = createZepProcessors({
2
  client,
3
  userId,
4
  threadId,
5
  // Replace thread.getUserContext with your own retrieval:
6
  contextBuilder: async ({ client, userId, threadId, userMessage }) => {
7
    const result = await client.graph.search({ userId, query: userMessage, scope: "edges" });
8
    return result.edges?.map((e) => e.fact).join("\n");
9
  },
10
  // Or just customize the wrapping template (must contain a literal `{context}`):
11
  contextTemplate: "Known facts about the user:\n{context}",
12
  // Or fully take over formatting (wins over contextTemplate):
13
  formatContext: (context) => `<memory>${context}</memory>`,
14
});

• contextBuilder replaces the default retrieval with your own async function. It receives a ZepContextBuilderInput — the client, the resolved userId/threadId, and the latest user message — and returns the context string (or undefined to inject nothing for that turn). The result still passes through the template or formatContext.
• contextTemplate customizes the wrapping text. It must contain a literal {context} placeholder, replaced via literal string replacement (not a format string), so braces, %, and $ in the retrieved context are safe.
• formatContext takes over formatting entirely and wins over contextTemplate.

Per-call identity

Pass resolveIdentity (a ZepIdentityResolver, sync or async) to resolve userId/threadId per call from Mastra’s requestContext instead of binding a fixed identity at construction time — useful when a single processor instance serves many end users:

1
const { inputProcessor, outputProcessor } = createZepProcessors({
2
  client,
3
  resolveIdentity: (requestContext) => ({
4
    userId: (requestContext as { userId?: string } | undefined)?.userId,
5
    threadId: (requestContext as { threadId?: string } | undefined)?.threadId,
6
  }),
7
});

The same resolveIdentity option is accepted by createZepSearchTool, createZepRememberTool, and createZepContextTool (resolved from each tool call’s context.requestContext), and createZepToolset forwards it to all three tools.

If both a fixed userId/threadId and resolveIdentity are set, resolveIdentity’s result wins for whichever fields it returns; any field it omits or resolves to undefined falls back to the constructor-bound value.

Provisioning with ensureZepUserAndThread

Zep requires the user and thread to exist before messages are added. Call ensureZepUserAndThread once, out-of-band, before the first turn. It creates then catches the conflict, so calling it repeatedly for the same user and thread is safe. Already-exists detection is typed — a 409 status, or a 400 with “already exists” — so genuine failures (auth, network, 5xx) are never mistaken for a conflict; they are logged at warn and reported via a false return, never thrown.

Pass onUserCreated (a ZepUserCreatedHook) to run one-time setup — per-user ontology, custom instructions, seeding — exactly once, only when the user is genuinely newly created:

1
await ensureZepUserAndThread({
2
  client,
3
  userId,
4
  threadId,
5
  firstName: "Jane",
6
  lastName: "Smith",
7
  email: "[email protected]",
8
  // Fires exactly once, only when the user is genuinely newly created —
9
  // e.g. configure per-user summary instructions:
10
  onUserCreated: async (client, userId) => {
11
    await client.user.addUserSummaryInstructions({
12
      userIds: [userId],
13
      instructions: [{ name: "diet", text: "Track the user's dietary preferences." }],
14
    });
15
  },
16
});

Tools

The toolset puts the model in the loop: the agent calls a tool when it decides to persist or recall. Use it standalone or alongside the processors.

1
import { ZepClient } from "@getzep/zep-cloud";
2
import { Agent } from "@mastra/core/agent";
3
import { createZepToolset, ensureZepUserAndThread } from "@getzep/zep-mastra";
4
5
const client = new ZepClient({ apiKey: process.env.ZEP_API_KEY! });
6
7
// 1. Provision the Zep user + thread before the first turn.
8
const binding = { userId: "user-123", threadId: "thread-abc" };
9
await ensureZepUserAndThread({ client, ...binding, firstName: "Jane", lastName: "Smith" });
10
11
// 2. Build the tool set bound to that user + thread.
12
const { zepRemember, zepSearch, zepContext } = createZepToolset({ client, binding });
13
14
// 3. Attach the tools to an Agent (id and name are both required by Mastra).
15
const agent = new Agent({
16
  id: "memory-agent",
17
  name: "Memory Agent",
18
  instructions: "You have long-term memory. Store and recall user facts.",
19
  model: "openai/gpt-5.6-terra",
20
  tools: { zepRemember, zepSearch, zepContext },
21
});

The toolset provides three tools:

Each tool is also exported as a standalone factory (createZepRememberTool, createZepSearchTool, createZepContextTool) for wiring a single tool with custom options.

Each tool has a typed input and output schema:

zepSearch returns facts as extracted strings tailored to the search scope — edge facts, "name: summary" for entities, episode content, and so on — with found set to true when the result is non-empty.

Pin-or-expose search parameters

createZepSearchTool exposes each graph.search parameter to the model by default, alongside the always-required query, so the model can tune its own searches per call:

Each parameter is independently tri-state at construction time (ZepSearchPinnableParams):

• pinnedParams fixes a parameter to a constant value: hidden from the model’s schema, always sent.
• hiddenParams removes a parameter from the schema without pinning it: omitted from the graph.search call entirely, so Zep’s own server default applies.
• Omitted from both — exposed to the model with the documented default.

1
// Model only ever sees `query`; scope/reranker/limit are fixed.
2
createZepSearchTool({
3
  client,
4
  binding: { userId },
5
  pinnedParams: { scope: "edges", reranker: "rrf", limit: 10 },
6
});
7
8
// Hide mmrLambda/centerNodeUuid from the schema without fixing a value.
9
createZepSearchTool({
10
  client,
11
  binding: { userId },
12
  hiddenParams: new Set(["mmrLambda", "centerNodeUuid"]),
13
});

searchFilters and bfsOriginNodeUuids are always constructor-only — never exposed to the model — and applied whenever set. The legacy scope/reranker/limit constructor arguments pin (and hide) their parameter, equivalent to the corresponding pinnedParams entry.

Binding: user graph vs standalone graph

Tools and processors are bound to a graph via userId/graphId (tools take these on a ZepBinding; the processors take userId/threadId directly):

• userId targets a user graph — the home for personalized agent memory. Use it for a conversational agent that remembers an end user. Context retrieval and the zepContext tool also need a threadId (the thread scopes relevance; retrieval still spans the whole user graph).
• graphId targets a standalone graph — shared or domain knowledge such as a product knowledge base or runbooks. No user node, no user summary. Standalone graphs are supported by the tools; the processors are thread-oriented and expect a userId.

If both are set, userId wins. If neither is set (or threadId can’t be resolved), tools and processors degrade gracefully instead of throwing.

Roles

zepRemember accepts an arbitrary role string and maps it onto Zep’s closed RoleType enum: user, assistant, system, tool, function, or norole. Host-framework role names like human or ai are coerced safely; unknown roles fall back to norole. The mapper is exported as toRoleType.

Best practices

• Use tools for untrusted context. Use processors only when all stored content is application-authored and trusted.
• Call ensureZepUserAndThread once before the first turn, then reuse a single ZepClient
• Pass real names so Zep can anchor the user’s identity node in the graph
• Don’t read-after-write within a turn — Zep builds the graph asynchronously, so a just-stored fact is not instantly retrievable
• Pass a custom logger to route Zep warnings into your logging stack

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