Crate kyma_memory 

Agentic Memory substrate for Kyma.

Memory is stored as ordinary Kyma columnar tables (memory_nodes / memory_edges) registered as the memory graph, so it is first-class queryable (run_sql/run_kql/Discover) and renders in the unified GraphView. Embeddings are a FixedSizeList<Float32> column searched via the platform cosine_distance UDF. Storage is append-only; mutations write a new version and recall dedups to the latest by updated_at.

See docs/superpowers/specs/2026-05-31-agentic-memory-design.md (M1).

Re-exports

pub use embed::build_embedding_backend;

pub use embed::shared_embedding;

pub use error::MemoryError;

pub use error::Result;

pub use types::CreateMemory;

pub use types::MemoryStatus;

pub use types::MemoryType;

pub use types::RecallFilter;

Modules

embed

Embedding-backend construction. Memory is the first server-side consumer of kyma-embed; the backend (an ONNX model for fastembed) is built once and shared process-wide via a OnceCell so the model loads at most once.

error

Error type for the memory layer.

rows

Build JSON rows for the columnar memory tables. Rows are serialized to NDJSON and coerced to Arrow by kyma_ingest_core::parse_ndjson, so the embedding value is a plain JSON array of floats.

schema

Arrow schemas for the columnar memory tables.

sql

SQL builders for memory recall / listing. The queries run server-side (where KymaTable + the cosine_distance UDF live) against the columnar memory tables. Each query first dedups to the latest version per node id (row_number() ... ORDER BY updated_at DESC), matching the append-only “latest-wins” model.

types

Memory domain types.

Structs

MemoryWriter

Writes memory data and provisions the memory database/tables/graph on demand.

Constants

DEFAULT_DATABASE

Dedicated database that holds the memory tables.

DEFAULT_REALM

Default per-item realm when none is supplied.

EDGE_DERIVED_FROM

EDGE_INVALIDATES

EDGE_MERGED_INTO

EDGE_REFERENCES

Canonical memory edge-relationship types.

EDGE_RELATES_TO

EDGE_RESOLVES_TO

EDGE_TABLE

GLOBAL_REALM

Shared realm for cross-project facts; recall unions project ∪ global.

GRAPH_NAME

Registered graph name for memory.

HALF_LIFE_DAYS

Half-life (days) for recency decay exp(-ln2 * age_days / HALF_LIFE_DAYS).

IMPORTANCE_WEIGHT

NODE_TABLE

RELEVANCE_WEIGHT

Recall re-rank blend: score = RELEVANCE_WEIGHT*(1-distance) + IMPORTANCE_WEIGHT*importance.

RRF_K

Reciprocal-rank-fusion constant in 1/(RRF_K + rank).

W_GRAPH

W_IMPORTANCE

W_KEYWORD

W_RECENCY

W_RRF

Graph-aware hybrid recall blend. The final score fuses reciprocal-rank fusion (RRF) of the vector + keyword candidate lists with semantic similarity, keyword match, graph proximity, importance, and recency (see the memory_retrieve orchestrator). Tunable; sane defaults below.

W_SEMANTIC