File size: 13,120 Bytes
0ae3f27
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
---
title: Graph Memory
description: "Layer relationships onto Mem0 search so agents remember who did what, when, and with whom."
icon: "network-wired"
---

Graph Memory extends Mem0 by persisting nodes and edges alongside embeddings, so recalls stitch together people, places, and events instead of just keywords.

<Info icon="sparkles">
**You’ll use this when…**
- Conversation history mixes multiple actors and objects that vectors alone blur together
- Compliance or auditing demands a graph of who said what and when
- Agent teams need shared context without duplicating every memory in each run
</Info>

## How Graph Memory Maps Context

Mem0 extracts entities and relationships from every memory write, stores embeddings in your vector database, and mirrors relationships in a graph backend. On retrieval, vector search narrows candidates while the graph returns related context alongside the results.

```mermaid
graph LR
    A[Conversation] --> B(Extraction LLM)
    B --> C[Vector Store]
    B --> D[Graph Store]
    E[Query] --> C
    C --> F[Candidate Memories]
    F --> D
    D --> G[Contextual Recall]
```

## How It Works

<Steps>
<Step title="Extract people, places, and facts">
Mem0’s extraction LLM identifies entities, relationships, and timestamps from the conversation payload you send to `memory.add`.
</Step>
<Step title="Store vectors and edges together">
Embeddings land in your configured vector database while nodes and edges flow into a graph backend (Neo4j, Memgraph, Neptune, Kuzu, or Apache AGE).
</Step>
<Step title="Expose graph context at search time">
`memory.search` performs vector similarity (optionally reranked by your configured reranker) and returns the results list. Graph Memory runs in parallel and adds related entities in the `relations` array—it does not reorder the vector hits automatically.
</Step>
</Steps>

## Quickstart (Neo4j Aura)

<Info icon="clock">
**Time to implement:** ~10 minutes · **Prerequisites:** Python 3.10+, Node.js 18+, Neo4j Aura DB (free tier)
</Info>

Provision a free [Neo4j Aura](https://neo4j.com/product/auradb/) instance, copy the Bolt URI, username, and password, then follow the language tab that matches your stack.

<Tabs>
  <Tab title="Python">
<Steps>
<Step title="Install Mem0 with graph extras">
```bash
pip install "mem0ai[graph]"
```
</Step>
<Step title="Export Neo4j credentials">
```bash
export NEO4J_URL="neo4j+s://<your-instance>.databases.neo4j.io"
export NEO4J_USERNAME="neo4j"
export NEO4J_PASSWORD="your-password"
```
</Step>
<Step title="Add and recall a relationship">
```python
import os
from mem0 import Memory

config = {
    "graph_store": {
        "provider": "neo4j",
        "config": {
            "url": os.environ["NEO4J_URL"],
            "username": os.environ["NEO4J_USERNAME"],
            "password": os.environ["NEO4J_PASSWORD"],
            "database": "neo4j",
        }
    }
}

memory = Memory.from_config(config)

conversation = [
    {"role": "user", "content": "Alice met Bob at GraphConf 2025 in San Francisco."},
    {"role": "assistant", "content": "Great! Logging that connection."},
]

memory.add(conversation, user_id="demo-user")

results = memory.search(
    "Who did Alice meet at GraphConf?",
    user_id="demo-user",
    limit=3,
    rerank=True,
)

for hit in results["results"]:
    print(hit["memory"])
```
</Step>
</Steps>
  </Tab>
  <Tab title="TypeScript">
<Steps>
<Step title="Install the OSS SDK">
```bash
npm install mem0ai
```
</Step>
<Step title="Load Neo4j credentials">
```bash
export NEO4J_URL="neo4j+s://<your-instance>.databases.neo4j.io"
export NEO4J_USERNAME="neo4j"
export NEO4J_PASSWORD="your-password"
```
</Step>
<Step title="Enable graph memory and query it">
```typescript
import { Memory } from "mem0ai/oss";

const config = {
  enableGraph: true,
  graphStore: {
    provider: "neo4j",
    config: {
      url: process.env.NEO4J_URL!,
      username: process.env.NEO4J_USERNAME!,
      password: process.env.NEO4J_PASSWORD!,
      database: "neo4j",
    },
  },
};

const memory = new Memory(config);

const conversation = [
  { role: "user", content: "Alice met Bob at GraphConf 2025 in San Francisco." },
  { role: "assistant", content: "Great! Logging that connection." },
];

await memory.add(conversation, { userId: "demo-user" });

const results = await memory.search(
  "Who did Alice meet at GraphConf?",
  { userId: "demo-user", limit: 3, rerank: true }
);

results.results.forEach((hit) => {
  console.log(hit.memory);
});
```
</Step>
</Steps>
  </Tab>
</Tabs>

<Info icon="check">
Expect to see **Alice met Bob at GraphConf 2025** in the output. In Neo4j Browser run `MATCH (p:Person)-[r]->(q:Person) RETURN p,r,q LIMIT 5;` to confirm the edge exists.
</Info>

<Note>
Graph Memory enriches responses by adding related entities in the `relations` key. The ordering of `results` always comes from vector search (plus any reranker you configure); graph edges do not reorder those hits automatically.
</Note>

## Operate Graph Memory Day-to-Day

<AccordionGroup>
  <Accordion title="Refine extraction prompts">
    Guide which relationships become nodes and edges.

<CodeGroup>
```python Python
import os
from mem0 import Memory

config = {
    "graph_store": {
        "provider": "neo4j",
        "config": {
            "url": os.environ["NEO4J_URL"],
            "username": os.environ["NEO4J_USERNAME"],
            "password": os.environ["NEO4J_PASSWORD"],
        },
        "custom_prompt": "Please only capture people, organisations, and project links.",
    }
}

memory = Memory.from_config(config_dict=config)
```

```typescript TypeScript
import { Memory } from "mem0ai/oss";

const config = {
  enableGraph: true,
  graphStore: {
    provider: "neo4j",
    config: {
      url: process.env.NEO4J_URL!,
      username: process.env.NEO4J_USERNAME!,
      password: process.env.NEO4J_PASSWORD!,
    },
    customPrompt: "Please only capture people, organisations, and project links.",
  }
};

const memory = new Memory(config);
```
</CodeGroup>
  </Accordion>
  <Accordion title="Raise the confidence threshold">
    Keep noisy edges out of the graph by demanding higher extraction confidence.

```python
config["graph_store"]["config"]["threshold"] = 0.75
```
  </Accordion>
  <Accordion title="Toggle graph writes per request">
    Disable graph writes or reads when you only want vector behaviour.

```python
memory.add(messages, user_id="demo-user", enable_graph=False)
results = memory.search("marketing partners", user_id="demo-user", enable_graph=False)
```
  </Accordion>
  <Accordion title="Organize multi-agent graphs">
    Separate or share context across agents and sessions with `user_id`, `agent_id`, and `run_id`.

<CodeGroup>
```typescript TypeScript
memory.add("I prefer Italian cuisine", { userId: "bob", agentId: "food-assistant" });
memory.add("I'm allergic to peanuts", { userId: "bob", agentId: "health-assistant" });
memory.add("I live in Seattle", { userId: "bob" });

const food = await memory.search("What food do I like?", { userId: "bob", agentId: "food-assistant" });
const allergies = await memory.search("What are my allergies?", { userId: "bob", agentId: "health-assistant" });
const location = await memory.search("Where do I live?", { userId: "bob" });
```
</CodeGroup>
  </Accordion>
</AccordionGroup>

<Note>
Monitor graph growth, especially on free tiers, by periodically cleaning dormant nodes: `MATCH (n) WHERE n.lastSeen < date() - duration('P90D') DETACH DELETE n`.
</Note>

## Troubleshooting

<AccordionGroup>
  <Accordion title="Neo4j connection refused">
    Confirm Bolt connectivity is enabled, credentials match Aura, and your IP is allow-listed. Retry after confirming the URI format is `neo4j+s://...`.
  </Accordion>
  <Accordion title="Neptune Analytics rejects requests">
    Ensure the graph identifier matches the vector dimension used by your embedder and that the IAM role allows `neptune-graph:*DataViaQuery` actions.
  </Accordion>
  <Accordion title="Graph store outage fallback">
    Catch the provider error and retry with `enable_graph=False` so vector-only search keeps serving responses while the graph backend recovers.
  </Accordion>
</AccordionGroup>

## Decision Points

- Select the graph store that fits your deployment (managed Aura vs. self-hosted Neo4j vs. AWS Neptune vs. local Kuzu vs. Apache AGE on PostgreSQL).
- Decide when to enable graph writes per request; routine conversations may stay vector-only to save latency.
- Set a policy for pruning stale relationships so your graph stays fast and affordable.

## Provider setup

Choose your backend and expand the matching panel for configuration details and links.

<AccordionGroup>
  <Accordion title="Neo4j Aura or self-hosted">
    Install the APOC plugin for self-hosted deployments, then configure Mem0:

```typescript
import { Memory } from "mem0ai/oss";

const config = {
    enableGraph: true,
    graphStore: {
        provider: "neo4j",
        config: {
            url: "neo4j+s://<HOST>",
            username: "neo4j",
            password: "<PASSWORD>",
        }
    }
};

const memory = new Memory(config);
```

Additional docs: [Neo4j Aura Quickstart](https://neo4j.com/docs/aura/), [APOC installation](https://neo4j.com/docs/apoc/current/installation/).
  </Accordion>
  <Accordion title="Memgraph (Docker)">
    Run Memgraph Mage locally with schema introspection enabled:

```bash
docker run -p 7687:7687 memgraph/memgraph-mage:latest --schema-info-enabled=True
```

Then point Mem0 at the instance:

```python
from mem0 import Memory

config = {
    "graph_store": {
        "provider": "memgraph",
        "config": {
            "url": "bolt://localhost:7687",
            "username": "memgraph",
            "password": "your-password",
        },
    },
}

m = Memory.from_config(config_dict=config)
```

Learn more: [Memgraph Docs](https://memgraph.com/docs).
  </Accordion>
  <Accordion title="Amazon Neptune Analytics">
    Match vector dimensions between Neptune and your embedder, enable public connectivity (if needed), and grant IAM permissions:

```python
from mem0 import Memory

config = {
    "graph_store": {
        "provider": "neptune",
        "config": {
            "endpoint": "neptune-graph://<GRAPH_ID>",
        },
    },
}

m = Memory.from_config(config_dict=config)
```

Reference: [Neptune Analytics Guide](https://docs.aws.amazon.com/neptune/latest/analytics/).
  </Accordion>
  <Accordion title="Amazon Neptune DB (with external vectors)">
    Create a Neptune cluster, enable the public endpoint if you operate outside the VPC, and point Mem0 at the host:

```python
from mem0 import Memory

config = {
    "graph_store": {
        "provider": "neptunedb",
        "config": {
            "collection_name": "<VECTOR_COLLECTION_NAME>",
            "endpoint": "neptune-graph://<HOST_ENDPOINT>",
        },
    },
}

m = Memory.from_config(config_dict=config)
```

Reference: [Accessing Data in Neptune DB](https://docs.aws.amazon.com/neptune/latest/userguide/).
  </Accordion>
  <Accordion title="Kuzu (embedded)">
    Kuzu runs in-process, so supply a path (or `:memory:`) for the database file:

```python
config = {
    "graph_store": {
        "provider": "kuzu",
        "config": {
            "db": "/tmp/mem0-example.kuzu"
        }
    }
}
```

Kuzu will clear its state when using `:memory:` once the process exits. See the [Kuzu documentation](https://kuzudb.com/docs/) for advanced settings.
  </Accordion>
  <Accordion title="Apache AGE (PostgreSQL extension)">
    [Apache AGE](https://age.apache.org/) adds graph database capabilities to PostgreSQL, letting you run Cypher queries alongside SQL on the same server. Start AGE via Docker, then configure Mem0:

```bash
docker run --name age-postgres \
  -e POSTGRES_DB=mem0_db \
  -e POSTGRES_USER=mem0_user \
  -e POSTGRES_PASSWORD=mem0_pass \
  -p 5432:5432 \
  -d apache/age
```

```python
from mem0 import Memory

config = {
    "graph_store": {
        "provider": "apache_age",
        "config": {
            "host": "localhost",
            "port": 5432,
            "database": "mem0_db",
            "username": "mem0_user",
            "password": "mem0_pass",
            "graph_name": "mem0_graph",
        },
    },
}

m = Memory.from_config(config_dict=config)
```

Apache AGE does not have a built-in vector index, so similarity search is computed client-side. This works well for moderate graph sizes; for very large graphs consider pairing AGE with pgvector for the vector store.

Reference: [Apache AGE documentation](https://age.apache.org/age-manual/master/index.html).
  </Accordion>
</AccordionGroup>

<CardGroup cols={2}>
  <Card
    title="Enhanced Metadata Filtering"
    description="Blend field-level filters with graph context to zero in on the right memories."
    icon="funnel"
    href="/open-source/features/metadata-filtering"
  />
  <Card
    title="Reranker-Enhanced Search"
    description="Layer rerankers on top of vectors and graphs for the cleanest results."
    icon="sparkles"
    href="/open-source/features/reranker-search"
  />
</CardGroup>