Tablestore:Product overview

Serverless and ready to use

Memory Storage uses a fully managed serverless architecture. Write and retrieve memories through APIs without deploying or maintaining infrastructure. Pay only for what you use, with no charges at zero usage.

Automatic memory extraction

After you write conversations or text, the service extracts key facts and preferences into structured memories. No custom extraction or organization logic is required. Use memory-store-level extraction instructions (extractInstructions) to focus extraction on business-critical information.

File view access

Service-managed structured memories support two read modes: structured access through SDKs, the CLI, and other interfaces, or directory- and file-based access through file views.

File views expose memory data as read-only files. Applications continue to write or update source memories through structured interfaces, preventing conflicts between file operations and structured memories.

File memory

If you already have Markdown or other UTF-8 text files, or want to organize a memory directory yourself, write the files directly to a file memory store. Applications control file paths and can create, read, update, rename, and delete files, as well as query file version history.

Automatic memory consolidation (Dream)

Memory consolidation (Dream) asynchronously refines stored memories in the background. It can merge, deduplicate, and update long-term memories (memory), and extract reusable skills (skill) and structured user profiles (profile) from interaction history. Dream can generate proposals for manual confirmation or automatically apply changes that meet a confidence threshold.

Horizontal scaling

Built on Tablestore's distributed architecture, Memory Storage scales horizontally. Storage capacity theoretically has no upper limit and can support millions of tenants and tens of billions of memories.

Accurate semantic retrieval

Memory Storage combines vector search and full-text search in a hybrid retrieval mode. Natural language queries recall relevant memories. Optionally enable Rerank, similarity threshold filtering (minSimilarity), and source evidence return (includeEvidence) to improve relevance and explainability.

Lower token costs

Compared with injecting full conversation histories, Memory Storage saves about 95% of tokens with almost no loss in response quality. Retrieving only relevant memories reduces unrelated context. It also avoids O(N²) cumulative token growth from repeatedly injecting the full history and the "Lost in the Middle" performance degradation.

Multi-tenant isolation

A four-level Scope of appId, tenantId, agentId, and runId isolates data by application, tenant, agent, and session for millions of tenants.

Continuity across multi-turn conversations

Write historical messages to a memory store during long conversations. The service retains the original messages and extracts key facts. Later conversations retrieve only relevant memories on demand, avoiding full-history context injection.

Preference reuse across sessions

User preferences, habits, and constraints expressed across sessions are extracted into long-term memories. New sessions retrieve these memories through semantic search, maintaining consistent agent responses across different times and entry points.

Knowledge sharing across multiple agents

Using a wildcard for agentId in the four-level Scope lets multiple agents share the same user memories within a tenant. This avoids extracting and maintaining the same facts for each agent.

User profile accumulation

Continuously write user behavior, attributes, and historical facts to a memory store. The service automatically consolidates the data into structured profile units that recommendation, customer service, and marketing agents can retrieve.