Memory System

Linea's memory system is inspired by SuperMemory and targets ~45% feature parity. It provides persistent, structured memory for AI agents across workflow runs: enabling agents to remember facts, user preferences, and past events.

Design Principles

• Atomic facts: raw text is split into the smallest self-contained statements before storing
• Semantic deduplication: new facts semantically equivalent to existing ones supersede them rather than duplicating
• Hybrid retrieval: pgvector cosine similarity + keyword search, merged with score boosting
• Temporal grounding: eventDate is separate from createdAt so facts can reference when something happened
• User profiling: scope=user memories grouped by factType form a user's behavioral profile

Schema

memories: {
  id: uuid PK
  workspaceId: uuid FK → workspaces
  userId: uuid FK → users (nullable)
  threadId: text (nullable)         // agent conversation thread
  workflowId: uuid FK → workflows (nullable)
  scope: enum('thread','workflow','user')
  content: text                     // the atomic fact
  embedding: vector(1536)           // OpenAI text-embedding-3-small
  source: enum('manual','extracted','ingested')  default: 'manual'
  factType: enum('fact','preference','event','profile','system') nullable
  eventDate: timestamp (nullable)   // when the fact occurred (≠ createdAt)
  supersededById: uuid (nullable)   // points to the replacement memory
  confidence: real                  // 0.0–1.0, default 1.0
  metadata: jsonb                   // arbitrary structured data
  createdAt: timestamp
  updatedAt: timestamp
}
 
memory_sessions: {
  id: uuid PK
  workspaceId: uuid FK → workspaces
  rawContent: text                  // original input text
  memoriesExtracted: text[]         // IDs of newly created memories
  memoriesUpdated: text[]           // IDs of superseded memories
  createdAt: timestamp
}

Ingest Pipeline

POST /workspaces/:id/memories/ingest

[
  { "content": "User prefers dark mode", "factType": "preference", "confidence": 0.97 },
  { "content": "User's name is Alex", "factType": "profile", "confidence": 0.99 },
  { "content": "Deployed to production on 2024-03-15", "factType": "event",
    "confidence": 0.95, "eventDate": "2024-03-15" }
]

SELECT id, 1 - (embedding <=> $1::vector) AS similarity
FROM memories
WHERE workspace_id = $2 AND superseded_by_id IS NULL
ORDER BY embedding <=> $1::vector
LIMIT 1

Response

{
  "memoriesCreated": 3,
  "memoriesUpdated": 1,
  "sessionId": "uuid",
  "memories": [...]
}

Hybrid Search

POST /workspaces/:id/memories/search

{
  "query": "what does the user prefer?",
  "scope": "user",
  "limit": 10
}

Algorithm:

1. Embed the query with text-embedding-3-small
2. pgvector cosine → top 20 candidates with scores
3. ilike keyword search → top 20 candidates
4. Merge scores:
   • Vector-only result: use raw cosine similarity
   • Keyword match on top of vector: score + 0.2
   • Keyword-only (not in vector results): score = 0.3
5. Exclude supersededById IS NOT NULL
6. Return top-K sorted by merged score

User Profile

GET /workspaces/:id/memories/profile?userId=...

Returns scope=user memories (excluding superseded) grouped by factType:

{
  "facts": [...],
  "preferences": [...],
  "events": [...],
  "profile": [...],
  "system": [...],
  "uncategorized": [...]
}

Services

EmbeddingService

Wraps OpenAI text-embedding-3-small. Zero-vector fallback if OPENAI_API_KEY is not set: useful for local dev without AI billing.

ExtractionService

Wraps Claude Haiku with a strict extraction prompt. Falls back to storing the entire input as a single fact if:

• ANTHROPIC_API_KEY is not set
• Claude returns non-JSON
• Any parse error occurs

Feature Parity vs SuperMemory