Recall

Recall uses semantic search combined with temporal filtering and graph traversal to find relevant memories.

Data Flow

Query Text
    │
    ▼
[1. Generate Embedding]         → 1536-dim vector (OpenAI)
    │
    ▼
[2. Vector Similarity Search]   → Qdrant, returns (MemoryID, score)
    │
    ▼
[3. Fetch Full Memories]        → XTDB
    │
    ▼
[4. Apply Filters]              → layer, entity, temporal (as_of)
    │
    ▼
[5. Temporal Re-ranking]        → superseded memories get 30% penalty
    │
    ▼
[6. Fetch Relationships]        → returned as metadata
    │
    ▼
RecallResponse { memories, usage }

Recall Modes

Mode	Context Returned	Use Case
SemanticSearch	None	Fast, simple queries
ContextualSearch	Parent + children + related	Understanding context
TemporalSearch	Entity history	Time-based queries

Query Parameters

Parameter	Type	Required	Description
query	string	Yes	Natural language search query
limit	int	No	Maximum results (default: 5)
layer	int (0-3)	No	Filter by semantic layer
as_of	ISO 8601	No	Query as of a specific point in time
expand_graph	bool	No	Include relationships and ancestors

Temporal Re-ranking

Superseded memories (those with valid_to set) receive a 30% score penalty:

adjusted_score = original_score × 0.7   (superseded)
adjusted_score = original_score × 1.0   (current)

Current information ranks higher, but historical data still surfaces if highly relevant.

API

curl "https://api.memlayer.dev/api/v1/recall?query=programming+language&limit=5&expand_graph=true"

Response includes ranked memories with similarity scores and optional graph context (relationships, ancestors).