Korely MCP

korely_search follows the rules of the REST search for types and full_text (Facts and whole text): with the facts, min_score is refused, and a memory comes first at equal rank. With full_text=true an answer holds 8,000 tokens of whole texts at most, the size of the largest korely_get_context block: past that, a result shows what is left (never less than its first 280 characters), its line says [cut], and a NOTE closes the answer. REST's content has no such bound.

Line breaks in a text. korely_search, korely_get_facts and korely_add (the facts it extracted) print one item a line. Every line of a memory's or a fact's text after its first is indented by four spaces, with or without types and full_text, so a line at the margin is always an item of its own: a text cannot add a line that looks like a fact, a score or a NOTE. Only a text with a line break (a line feed, a carriage return, a form feed, or Unicode's line and paragraph separators) reads differently from 1.0.12; the result's id stays at the end of its last line. REST answers JSON and does not change. Since 1.0.16 the context block (korely_get_context, GET /v1/context) follows the same rule for its memories and facts: before, it printed a memory's line breaks as they were.

The first four tools take an optional user_id to scope to one end user. The two write tools, korely_correct_fact and korely_forget_memory, need a key with the memories:write scope, as their REST twins do, and treat the ids of another project (or, for a confined key, of another end user) as not found. A client lists the tools again on its next session to see them. The three read tools, korely_get_context, korely_search and korely_get_facts, also take an optional agent_id, to read one agent's namespace, and korely_get_facts takes predicate (for example works_at), as their REST twins do. korely_add takes an optional agent_id and run_id too, as its REST twin does. Under a burst from one account, korely_search and korely_get_context wait their turn, up to 20 seconds for a slot on the account and up to 20 more for the embedder (about 40 seconds at worst). If the wait runs out, korely_search answers a tool error, search_unavailable, and korely_get_context answers the block it could build, with a note that a part could not be read (an error only when the block is empty). A tool refuses an argument it does not declare, as a REST body refuses an unknown field: it answers error: invalid_request, <name> is not an argument of <tool>; it takes: .... The tools map onto the same REST endpoints the SDK uses, one store, every surface.

Verify it

A quick reachability + auth check, the same key as a bearer:

curl https://api.korely.ai/v1/ping \
  -H "Authorization: Bearer kor_live_..."
# {"ok": true, "tier": "hobby", "region": "eu-hel1", "processing_region": "global", "scopes": [...]}

In your client, ask it to recall something for a user, then to remember a new fact, then recall again, it should call korely_get_context and korely_add on its own.

Local stdio server (alternative)

Prefer a local process over the hosted endpoint, for a client that only speaks stdio? The first four tools above ship as a stdio server, korely-mcp, in the Python package. It is a thin client, not an offline store: every tool call goes to the REST API at api.korely.ai with your key, so it needs the network like the hosted endpoint. Since 0.1.16 it needs no extra dependency (the [mcp] extra is kept, empty) and runs on Python 3.9 or later:

pip install 'korely-memory[mcp]'

claude mcp add korely -- korely-mcp

korely-mcp reads the key from ~/.korely/config.json (written by korely init) or KORELY_API_KEY. The same four tools and the same store, with fewer options: the local reading tools take no agent_id (korely_add does), no predicate and no exclude_sessions, and korely_search returns 10 hits by default. The local stdio server does not send types or full_text yet. Its korely_get_context and korely_get_facts take known_as_of from version 0.1.22. On the remote server both tools take reader since 1.0.15 (see Agent notices); the local server does not send it yet. See Vibecoding for the full coding-assistant walkthrough.

What's next

• Wire the memory into your own agent: the Quickstart and SDK reference.
• Understand the moat: temporal facts and the graph.