What kyma is

The context engine for coding agents. One place your agent recalls durable memory, queries live data (logs, traces, code, data sources) in KQL or SQL, and walks the graph that ties a decision to the resources it's about.

A memory store remembers what you told it. A context engine also knows your live systems and how everything connects — and hands all of it to your agent through one surface (a Claude Code plugin, a CLI, or MCP).

The three layers

kyma is one engine that does three things, in order of how you'll meet them:

• Memory — durable, graph-aware recall across sessions and machines. Capture → LLM extraction with bi-temporal conflict resolution → near-realtime hybrid recall (vector + keyword + graph). The "it remembers" layer your agent feels first.
• Live data — every signal your stack emits (logs, traces, metrics, tool calls, prompt/response bodies, deploy events, config diffs) ingested through OTLP, REST, Kafka, or file-drop, plus scheduled data sources — answered in KQL or SQL at sub-second latency over years of history.
• The graph — entities and relationships as a first-class layer. Memories resolve to the real nodes they're about (a repo, a service, a table, a trace), so recall comes back wired to the systems it concerns.

The engine underneath

The reason "live data" means a decade of history answered in milliseconds — not a toy key-value store:

• Columnar Arrow on object storage you own (S3, MinIO, any object_store impl), with per-extent column statistics and token indices that make 99 %+ of queries skip 99 %+ of data.
• KQL, SQL, or PromQL over Arrow Flight gRPC — exact rows, streamed zero-copy — so an agent can ask twenty exploratory questions per prompt without melting a credit card.
• Federates with your operational databases — Postgres, MySQL, MongoDB register as catalogs and DataFusion joins them with kyma's own tables in one query.
• Scales from one binary to many nodes without a rewrite: the catalog is externalized from byte one, compute is stateless, object storage is the source of truth. See Local & cluster mode.

What it is not

• Not just a memory layer. Memory SaaS gives the agent facts; kyma gives it the facts, the live logs/traces/code those facts are about, and the graph linking them — one query surface, one MCP server.
• Not a metrics database. Metrics are one signal among many; treating them specially fragments the surface an agent needs to reason across.
• Not a log search appliance. Full-text on the dynamic column works, but the engine is column-aware first; a search-shaped query plans like a filter, not like a Lucene query.
• Not an OLAP warehouse for business data. Use it for memory, observability, and operational state — and for joining production telemetry with your customer table — not as your primary BI cube.

Who it's for

The coding agent that needs to remember, and to reason across the whole stack — and the developer tired of an agent that forgets every session and can't see production.

Where to go next

• The whole picture, end to end: How kyma works.
• Give your agent memory in five minutes: Quickstart.
• How recall and the graph work: Agentic Memory.
• How a query stays fast: The pruning cascade.