AnalyticDB:Long-term memory service

Memory scenarios

The service supports the following three memory scenarios:

Memory categories

For each memory, ADBPGMem automatically classifies the content and supports category-based retrieval. A single memory can have multiple category tags.

Memory retrieval

1. Query processing

   • Uses an LLM to extract key information from the user's original query.

   • Generates appropriate filtering conditions, such as categories or time ranges, based on context and query intent to improve retrieval efficiency and accuracy.

   • Configurable retrieval parameters: the number of memories to return (TopK) and a retrieval threshold (Threshold). Memories with scores below the threshold are not returned.

2. Vector search

   • Performs semantic search in a vector database by creating an embedding from the optimized query.

   • Ranks results based on their relevance to the query.

   • Applies specified filters (user, agent, metadata, etc.).

3. Result processing

   • Integrates results from multiple search criteria and ranks them by relevance, importance, and time.

   • Returns memory entries with their relevance scores.

   • Includes relevant metadata and timestamps, such as tags and categories.

Memory update

1. Information extraction

   • Uses an LLM to extract relevant memories from the conversation context.

   • Identifies important entities and extracts relationships between them to build a knowledge graph.

2. Memory storage

   • A vector database stores the actual memory content.

   • A graph database stores the relationships between entities.

   • Memory content is automatically updated in each turn of the conversation.

Hermes agent

Hermes uses a plugin architecture for its memory system (MemoryProvider ABC). The official built-in mem0 plugin communicates with the memory backend through the mem0 SDK. Any mem0-compatible service can serve directly as the memory backend for Hermes.

Mem0 SDK compatibility

ADBPGMem implements full compatibility with the mem0 SDK's /v3/ paths on top of its REST API.

Key adaptations:

• Path mapping: Maps /v3/memories/search/ to the internal semantic search engine.

• Parameter format compatibility: Accepts both a top-level user_id parameter and the nested filters: {user_id: ...} format.

• Response format alignment: The response structure is consistent with the mem0 Platform API ({results: [...], total, page, ...}).

No adapters or custom HTTP clients are required — any framework that uses MemoryClient(host=...) is plug-and-play, including mem0 ecosystem applications such as Hermes, LangChain, and CrewAI.

Hermes mem0 plugin changes

The mem0 plugin in Hermes originally supported only the official Mem0 Platform. To integrate with ADBPGMem, add a base_url configuration option to the plugin:

# plugins/memory/mem0/__init__.py (+9, -1)

# Core change in _get_client():
- self._client = MemoryClient(api_key=self._api_key)
+ kwargs = {"api_key": self._api_key}
+ if self._base_url:
+ kwargs["host"] = self._base_url
+ self._client = MemoryClient(**kwargs)

Configure ~/.hermes/.env:

MEM0_API_KEY=sk-your-adbpgmem-token
MEM0_BASE_URL=https://your-adbpgmem-endpoint

After Hermes starts, ADBPGMem automatically handles memory extraction, storage, and semantic retrieval on the server side.