Memory

The memory service provides semantic memory stores for AI agents — enabling them to remember context across conversations using vector-based storage and retrieval. Operators manage stores under Data → Agent Memory.

Agent Memory

A memory store is the unit that binds an embedding model to a vector index — it's where individual memory items live, scoped to a project. The list view shows total stores, total memories across all stores, and how many stores are currently active.

Creating the first store requires two prerequisites: a vector provider configured in Knowledge Index and an embedding model deployed in Model Hub. The Create First Store flow walks you through both — it'll halt and point to the missing piece if either isn't ready.

Once a store exists, every memory item carries content, free-form tags, an importance score, and an optional scope. Agents read from the store via the search() and recall() APIs documented below — recall() is the chat-optimised variant that respects a max-tokens budget.

Concepts

Concept	Description
Memory Store	Configuration container linking an embedding model and vector index
Memory Item	A single memory entry with content, tags, scope, and importance
Search	Semantic similarity search across stored memories
Recall	Chat-optimized retrieval with token limit awareness

Service Functions

Function	Description
createMemoryStore()	Create store with vector index + embedding model
listMemoryStores()	List stores for project
getMemoryStore()	Get store by key
updateMemoryStore()	Update config/status
deleteMemoryStore()	Delete store + vector index
addMemory()	Add single item (embed → store)
addMemoryBatch()	Batch add items
listMemoryItems()	List items in store
getMemoryItem()	Get single item
updateMemoryItem()	Update content (re-embed if changed)
deleteMemoryItem()	Delete single item
deleteMemoryItemsBulk()	Bulk delete
searchMemories()	Semantic search
recallForChat()	Chat-optimized recall with token limit

Create Store Request

{
  "name": "Agent Working Memory",
  "vectorProviderKey": "pinecone-prod",
  "embeddingModelKey": "text-embedding-ada-002"
}

The store key is generated automatically, and creating a store provisions the backing vector index.

Client API

Create Memory Store

POST /api/client/v1/memory/stores
Authorization: Bearer <token>

Add Memory

POST /api/client/v1/memory/stores/:storeKey/memories
Authorization: Bearer <token>

{
  "content": "User prefers dark mode and concise responses",
  "scope": "user",
  "tags": ["preferences", "ui"],
  "importance": 0.8,
  "metadata": { "userId": "user-123" }
}

Search Memories

POST /api/client/v1/memory/stores/:storeKey/search
Authorization: Bearer <token>

{
  "query": "What are the user's UI preferences?",
  "topK": 5,
  "scope": "user",
  "scopeId": "user-123",
  "tags": ["preferences"],
  "minScore": 0.7
}

Response:

{
  "memories": [
    {
      "id": "mem_123",
      "content": "User prefers dark mode and concise responses",
      "score": 0.94,
      "scope": "user",
      "scopeId": "user-123",
      "tags": ["preferences", "ui"],
      "metadata": { "userId": "user-123" },
      "importance": 0.8
    }
  ],
  "query": "What are the user's UI preferences?",
  "storeKey": "mem-agent-working-memory"
}

Recall for Chat

POST /api/client/v1/memory/stores/:storeKey/recall
Authorization: Bearer <token>

{
  "query": "What does this user like?",
  "scope": "user",
  "scopeId": "user-123",
  "topK": 5,
  "maxTokens": 500
}

The recall function is optimized for chat contexts. It returns a compact context string together with the matched memories, trimmed to fit the requested token budget.

Memory Properties

Field	Description
content	The memory text
scope	Categorization (e.g., user, session, global)
source	Origin identifier
tags	Array of string tags for filtering
importance	Float 0-1, influences ranking
metadata	Arbitrary key-value pairs

Pipeline

Adding a memory:

Content → Embed (via inference service) → Vector Upsert → Store metadata

Searching memories:

Query → Embed → Vector Search → Rank by score + importance → Return

Dependencies

• Inference Service — For generating embeddings
• Vector Service — For storing and querying vectors