Agent Memory Service (Private Preview)
Agent Memory Service (AMS) is a cloud-based long-term memory service for Agents that stores, retrieves, and manages memory across sessions. This allows an Agent to retain user preferences, project context, and historical decisions across days, devices, and instances.
AMS provides two types of entry points:
A Mem0-compatible API for applications and Agent plugins.
A control plane for platforms and consoles to manage tenants, users, workspaces, sessions, and API Keys.
This feature is currently in private preview. To request access, contact our technical staff in DingTalk group 59530004993 to have the feature enabled.
How AMS works
From write to recall, the AMS workflow consists of six steps:
AMS receives messages through the Mem0-compatible API and uses the API Key to identify the tenant, workspace, and session.
The service organizes continuous conversations into manageable interaction fragments based on time and semantic boundaries.
The memory engine extracts summaries, facts, participants, timestamps, and contextual cues from these fragments.
The system structures this information as long-term memory, user profiles, and event records, then writes it to storage and an index.
During retrieval, AMS combines keywords, semantic similarity, and a reranking strategy to find memories relevant to the current task.
The Agent uses the retrieved memory to answer, plan, or execute subsequent actions without replaying large amounts of raw historical text.
What problems does it solve?
Saves users from repeating their preferences, project background, and historical decisions.
Keeps long-term historical data out of the prompt.
Extracts facts, preferences, events, and user profiles from chat logs.
Isolates data in a production environment using tenants, workspaces, sessions, and API Keys.
Core capabilities
Capability | Description |
long-term memory construction | Extracts reusable memory from conversations and events. |
smart recall | Supports keyword, semantic, and hybrid recall. |
Mem0 compatibility | Allows integration with the Mem0 SDK, the OpenClaw plugin, or similar memory clients. |
multi-tenant governance | Manages tenants, users, workspaces, sessions, and API Keys. |
asynchronous write | Memory extraction runs as a background process to avoid blocking the main conversation flow. |
Core concepts
Concept | Description |
tenant | An isolation boundary for an organization or a customer. |
user | The entity that makes API calls or owns the memory. It can be a person, an Agent, or a service account. |
workspace | A dedicated space for memory. |
session | The context of a single conversation or thread, corresponding to the Mem0 |
API Key | The entry point for authentication and context resolution. |
memory entry | A memory item that can be externally retrieved, updated, and deleted. |
event | A processing event for a write request, used for querying asynchronous status. |
Use cases
Coding Agents that remember project structures, test commands, technical decisions, and user preferences.
Enterprise assistants that retain user or business context across devices and sessions.
Customer service Agents that provide continuous support based on past issues and their resolution progress.
Multi-agent systems that isolate memory by workspace and are managed by a unified control plane.
AMS is designed primarily to store information that will be reused. Short-term chat caches, pure log archives, or one-time Q&A interactions typically do not need to be written to AMS.