Building an AI agent with long memory application

Key capabilities

Memory retrieval

1. Query processing

The LLM extracts key information from the user's query and generates filter conditions (such as category or time range) to narrow the search.

2. Vector retrieval

The system creates an embedding from the refined query, runs a semantic search against the Vector Store, sorts results by relevance, and applies any specified filters (user, agent, or metadata).

3. Result processing

Results from multiple search conditions are merged and re-ranked by relevance, importance, and recency. Each result includes a relevance score, metadata, and a timestamp.

A single search call triggers this entire pipeline — no manual orchestration needed.

Memory update

1. Information extraction

The LLM analyzes the conversation context to extract relevant memories, identifies key entities, and maps the relationships between them for the knowledge graph.

2. Conflict resolution

The system compares new memories against existing ones. If a conflict is detected (same entity, different attribute), it resolves it automatically. The current strategy is timestamp priority — the most recent information wins.

Conflict resolution triggers when new information contradicts an existing memory entry in terms of entities, relationships, or attributes. The system uses three strategies:

3. Memory storage

The Vector Store saves the actual memory content. The Graph Store saves entity relationships. Memory content updates automatically with each conversation turn.

A single add call triggers this entire pipeline — no manual orchestration needed.

Memory types

AnalyticDB for PostgreSQL supports four memory types, modeled after cognitive science classifications:

Choose a memory scenario

Use the following table to decide which ID parameter to use when calling the memory interfaces.

Quick guide: Use run_id when you only need memories from the current session. Use user_id for personalization that spans sessions. Use agent_id when building agents that need to remember how they performed subtasks.

Configure the memory service

adbpg_llm_memory.config(config json)

Call this function to set up the LLM, embedding model, and Vector Store before running any other memory operations.

This configuration is session-scoped. All subsequent memory operations (add, search, get_all, delete_all) must run in the same session. If you're using DMS to test these examples, run all SQL statements in the same SQL Console tab.

Replace the api_key, user, password, and dbname values before running:

select adbpg_llm_memory.config(
    $$
    {
    "llm": {
        "provider": "qwen",
        "config": {
            "model": "qwen3-32b",
            "qwen_base_url": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
            "api_key": "sk-xxxxxxx"
        }
    },
    "embedder": {
        "provider": "openai",
        "config": {
            "model": "text-embedding-v3",
            "embedding_dims": "256",
            "api_key": "sk-xxxxxx",
            "openai_base_url": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
        }
    },
    "vector_store": {
      "provider": "adbpg",
      "config": {
            "user": "username",
            "password": "password",
            "dbname": "testdb",
            "hnsw": "True",
            "embedding_model_dims": "256"
            "port": 3029          
          }
      }
}
    $$
);

Expected output:

{"message": "Configuration set successfully"}

Add memories

adbpg_llm_memory.add(messages json, user_id text, run_id text, agent_id text, metadata json, memory_type text, prompt text)

The LLM extracts key facts from the conversation and stores each one as a separate memory entry.

SELECT adbpg_llm_memory.add($$
[
    {"role": "user", "content": "Hi, I'm Zhang San. I like hiking, but I don't like intense exercise,"},
    {"role": "assistant", "content": "Hello, Zhang San! Hiking is a great hobby. I'll remember your preferences. If you have any questions about hiking route planning, equipment recommendations, or scenic spots along the way, feel free to ask anytime."}
]
$$, 'test_u', null, null, $${"expiration_date": "2025-08-01"}$$, null, null);

Expected output:

{
    "results": [
        {
            "id": "e6d241f9-634f-43e4-925c-0ed70974****",
            "memory": "Name is Zhang San",
            "event": "ADD"
        },
        {
            "id": "9efbb099-a20b-483e-99ef-3cc1e85e****",
            "memory": "Likes hiking",
            "event": "ADD"
        },
        {
            "id": "6fc474d5-1e77-48ec-a5f2-8cb9ec50****",
            "memory": "Dislikes intense exercise",
            "event": "ADD"
        }
    ]
}

Get all memories for a user or agent

adbpg_llm_memory.get_all(user_id text, run_id text, agent_id text)

Returns all stored memories for the specified user or agent in JSON format.

SELECT adbpg_llm_memory.get_all('test_u', null, null);

Expected output:

{
    "results": [
        {
            "id": "1cf1e872-5f78-41d0-b1ab-370eee82****",
            "memory": "Name is Zhang San",
            "hash": "d6f327d1ea38b8387927810bdcd3****",
            "metadata": {"expiration_date": "2025-08-01"},
            "created_at": "2025-06-25T15:45:58.687949+08:00",
            "updated_at": null,
            "user_id": "test_u"
        },
        {
            "id": "5806ab99-9764-4f31-bbda-29b77e8b****",
            "memory": "Likes hiking",
            "hash": "5f8275169192f1a1a4564149c3d1****",
            "metadata": {"expiration_date": "2025-08-01"},
            "created_at": "2025-06-25T15:45:58.700653+08:00",
            "updated_at": null,
            "user_id": "test_u"
        },
        {
            "id": "55babe14-2605-40fa-9b32-e5ccefc3****",
            "memory": "Dislikes intense exercise",
            "hash": "18fa10d79b6d2b0ec7f271817095****",
            "metadata": {"expiration_date": "2025-08-01"},
            "created_at": "2025-06-25T15:45:58.704473+08:00",
            "updated_at": null,
            "user_id": "test_u"
        }
    ]
}

Search memories by query

adbpg_llm_memory.search(query text, user_id text, run_id text, agent_id text, filter json)

Retrieves memories semantically related to a given query. Use this to find relevant context before calling your LLM.

SELECT adbpg_llm_memory.search(
    'Can you recommend some exercise activities and locations for this weekend?',
    'test_u', null, null, null
);

Expected output — each result includes a score indicating semantic relevance:

{
    "results": [
        {
            "id": "5806ab99-9764-4f31-bbda-29b77e8b****",
            "memory": "Likes hiking",
            "hash": "5f8275169192f1a1a4564149c3d1****",
            "metadata": {"expiration_date": "2025-08-01"},
            "score": 0.4617832899093628,
            "created_at": "2025-06-25T15:45:58.700653+08:00",
            "updated_at": null,
            "user_id": "test_u"
        },
        {
            "id": "55babe14-2605-40fa-9b32-e5ccefc3****",
            "memory": "Dislikes intense exercise",
            "hash": "18fa10d79b6d2b0ec7f271817095****",
            "metadata": {"expiration_date": "2025-08-01"},
            "score": 0.48010164499282837,
            "created_at": "2025-06-25T15:45:58.704473+08:00",
            "updated_at": null,
            "user_id": "test_u"
        },
        {
            "id": "1cf1e872-5f78-41d0-b1ab-370eee82****",
            "memory": "Name is Zhang San",
            "hash": "d6f327d1ea38b8387927810bdcd3****",
            "metadata": {"expiration_date": "2025-08-01"},
            "score": 0.6027468977721387,
            "created_at": "2025-06-25T15:45:58.687949+08:00",
            "updated_at": null,
            "user_id": "test_u"
        }
    ]
}

Delete all memories for a user or agent

adbpg_llm_memory.delete_all(user_id text, run_id text, agent_id text)

Deletes all memories associated with the specified user, session, or agent.

SELECT adbpg_llm_memory.delete_all('test_u', null, null);

Expected output:

{"message": "All relevant memories deleted"}