Reference

MCP tools

Once connected, your agent sees the tools below. In practice you’ll use the four memory tools and episode_log constantly; the rest are there when you need them.

Memory

The core loop: persist a fact, recall facts, fetch one, forget one.

memory_add

Persist a single durable fact to long-term memory. Call it proactively whenever the user states a lasting fact, preference, or decision — don’t wait to be asked. For a multi-message exchange, prefer episode_log.

Field	Type	Required	Notes
`content`	string	yes	The fact to store.
`metadata`	object	no	Arbitrary key/value tags stored alongside the memory.

memory_search

Search memory with hybrid semantic + full-text ranking. Call it before answering anything that may depend on the user’s history, preferences, people, projects, or past decisions.

Field	Type	Required	Notes
`query`	string	yes	Natural-language query.
`limit`	number	no	Max results (default 10).
`kind`	string	no	Filter by memory kind (see Memory model).
`min_confidence`	number	no	Drop results below this score (0–1).

memory_get / memory_forget

Fetch a single memory by ID (memory_get), or soft-delete one (memory_forget). Both take a single memory_id string.

Episodes & context

episode_log

Persist a whole conversation with automatic memory formation: it stores the raw turns and extracts atomic, self-contained facts (names resolved, relative dates converted to absolute). Prefer this over memory_add for any multi-message context, and call it when a task concludes.

Field	Type	Required	Notes
`messages`	array	yes	Array of `{ role, content }` objects.
`source`	string	no	Optional label for where the episode came from.

context_assemble

Build a single, token-budgeted context bundle from the most relevant memories (and optionally skills and graph neighbours) for a query — one call instead of orchestrating several searches.

Field	Type	Required	Notes
`query`	string	yes	What the context is for.
`token_budget`	number	no	Cap on the assembled bundle size.

Knowledge graph

graph_query

Traverse the knowledge graph of entities and relationships extracted from your memories. Graph expansion also runs under the hood during memory_search to pull in related facts, so you get its benefit without calling it directly.

Skills

Skills are reusable, versioned instruction bundles your agents can list and load on demand: skill_list, skill_get (by slug + optional version), and skill_run_context (execute a skill and return its context block).

Configuration

mcp_config_get and mcp_config_list read the MCP client configurations saved for the current scope — useful for agents that provision other agents.

Every tool call is metered as one memory operation and counts toward your plan’s monthly quota. See the memory model for how scopes, kinds, and limits work.

Calling a tool without an agent

Everything above is also available over plain HTTP. For example, a search:

bash

curl -s "https://api.kognite.dev/api/memories?query=where%20do%20I%20deploy&limit=5" \
  -H "Authorization: Bearer YOUR_KEY"

See the REST API for the full endpoint list.