The kyma-memory Claude Code plugin

Install the plugin to turn Claude Code into a live source and consumer of Kyma's memory layer — it captures your session in realtime, auto-recalls relevant memories on each prompt, and distills learnings into durable memory at session end.

kyma connect https://your-kyma.example.com --token <token>
kyma install-plugin            # → ~/.claude/skills/kyma-memory

Restart Claude Code and run /kyma-status. install-plugin templates your server URL + token into the plugin's .mcp.json so both the hooks and the MCP server work out of the box. It lives in-repo at integrations/claude-code/kyma-memory/.

Note — engine vs. plugin: This is distinct from the Claude Code engine (Kyma using claude as an LLM backend). Here it's the other direction: Claude Code using Kyma as its memory and data backend.

What it does

Two lanes, one connection (it reuses your kyma connect credentials — the plugin stores none of its own):

Lane	Mechanism	Result
Firehose	hooks → kyma ingest push → POST /v1/ingest	Every meaningful turn (prompts, tool actions, assistant replies) lands in default.claude_code_events, queryable in KQL/SQL and visible in Discover/graph — in realtime.
Recall	UserPromptSubmit hook → kyma recall (MCP recall_memory)	Relevant durable memories are injected as context on each prompt.
Distill	SessionEnd hook + /kyma-remember → kyma distill → /v1/agent/ask	The kyma agent extracts durable memories (decisions, preferences, learnings) at session end.
Direct tools	bundled MCP server → /mcp/v1	The model can call recall_memory, save_memory, run_kql, run_sql, and the other Kyma tools natively.

Hooks

The plugin registers five lifecycle hooks (all fail open — capture never blocks a session):

Event	Action
SessionStart	firehose a session marker + inject a memory recap
UserPromptSubmit	firehose the prompt + inject recalled context
PostToolUse	firehose tool events (Edit/Write/Bash/Task/WebFetch/NotebookEdit)
Stop	firehose the assistant's final message
SessionEnd	distill the session into durable memories (detached)

Slash commands

• /kyma-recall <query> — semantic recall from durable memory.
• /kyma-remember [note] — save a memory, or distill the recent conversation.
• /kyma-ask <question> — ask Kyma about your data (logs/traces/tables/graph) in KQL/SQL.
• /kyma-ingest [source | resources] — pull from a data source on demand, or create virtual graph entities (ingest_entity) wired to memory and existing resources.
• /kyma-status — connection, capture mode, and this project's event breakdown.

The capture table

default.claude_code_events (auto-created, schema-evolving):

claude_code_events
| where realm == "kyma"
| summarize events = count() by kind
| order by events desc

kind is one of session_start · user_prompt · tool_use · assistant_turn · session_end. realm defaults to the project directory's basename.

Configuration & privacy

Behavior is controlled by environment variables (KYMA_CC_CAPTURE = off/metadata/full, KYMA_CC_AUTO_RECALL, KYMA_CC_DISTILL, KYMA_CC_REALM, …). In full mode conversation content is sent to your own Kyma server; common secret shapes are redacted before egress, and metadata mode captures event shape without bodies. See the plugin's README.md for the full table.

Uninstall

rm -rf ~/.claude/skills/kyma-memory