knowmind

knowmind is a managed AI memory service that provides persistent memory and a typed knowledge graph for AI tools. It connects to Claude Code, Cursor, and other MCP clients via a CLI and MCP server (local stdio), letting AI tools retrieve stored facts before answering instead of guessing.

Purpose knowmind is used to provide AI tools with persistent memory across sessions and clients. Once something is stored, it can be used in Claude Code, Cursor, and any other client running a local stdio MCP server. It is suited to individuals and teams that need an auditable memory service with data located in the EU.

Key capabilities

• Bi-temporal history: every fact carries two timelines—the time the fact was true and the time it was recorded. Corrections do not overwrite the original content; old statements remain queryable and retain their invalidation date, so the state of the system as known on any given day can be reconstructed.
• Source and confidence for every fact: new facts require a confidence level and retain source links; old facts without confidence are flagged.
• Typed knowledge graph: entities and relationships use a restricted set of edge types, and vague catch-all relationships are rejected at write time.
• Hybrid retrieval: combines keyword search (BM25), vector similarity, and the knowledge graph, then reranks with a cross-encoder.
• EU-operated, tenant-isolated: the service runs on Hetzner in Nuremberg, Germany, with tenant isolation, a public data processing agreement, and a subprocessor list (knowmind.de/legal/avv). Access logs use a hash chain with external timestamps (RFC 3161).
• The free tier is free forever, requires no payment information, and includes 2,500 memories, 1 user, and 30 days of access logs. The API, CLI, and MCP server are included in all tiers.

How to use Connect an AI tool in three steps:

1. Create a free account at knowmind.de and generate an access token: Dashboard → API tokens.
2. Store the token locally.
3. Configure the client:
   • Claude Code (one command).
   • Cursor (~/.cursor/mcp.json), which also applies to Claude Desktop, Windsurf, Cline, Continue, Goose, and Zed, using their respective configuration files.
   • Other MCP clients:knowmind is a standard stdio MCP server (npx -y knowmind mcp). On Windows, if npx cannot be launched directly, use "command": "cmd", "args": ["/c", "npx", "-y", "knowmind", "mcp"].
   • npx -y knowmind install <ide> will write the configuration for you (for example claude, cursor, vscode), and npx -y knowmind init sets up automatic memory hygiene for clients with a hook or rule mechanism (Claude Code, Cursor)—retrieve before each task, store after each meaningful change.

MCP tools (thirteen in total, all targeting your own workspace; the server does not see other tenants' data)

• knowmind_recall_memories: searches memories and returns passages that can answer the question, ranked. Parameters: query (string, required), k (integer, default 5, max 20), hops (integer, default 2, the distance to follow graph edges from the hit point). When there is no match, returns an empty result set and never fabricates an answer.
• knowmind_recall_at_time: the same search, but based on the state of memories on a specified date. Facts superseded after that date are excluded; facts valid at that time are returned even if they are no longer true now. Parameters: query (required), as_of (ISO date or timestamp, actually required), k (default 5).
• knowmind_list_recent: the most recently added entries, newest first. Parameters: k (default 10).
• knowmind_list_relations: all typed edges on an entry, in both directions. Parameters: memory_id (required). An unknown id returns an error, not an empty list.
• knowmind_get_schema: the entity classes and relation types accepted by this workspace. No parameters. Use it before calling upsert_entity or link_entities.
• knowmind_get_stats: corpus size: documents, chunks, entities, edges. No parameters.
• knowmind_get_health: whether the service and storage are reachable. No parameters.
• knowmind_store_memory: stores a short fact, decision, or note, kept as a single unit and not split. Parameters: content (required), title, memory_type (such as semantic, episodic), tags (string array), source, domain, relations (array, edges created along with the entry). When the text contains what appears to be a password or access key, returns 422 and rejects it.
• knowmind_upload_document: stores longer text as a document, splits it into paragraphs, embeds it, and makes it searchable via recall_memories. Parameters: content (required), title, source, relations (array).
• knowmind_update_fact: supersedes a fact that has changed. The old version retains its validity window and remains auditable; nothing is deleted. Parameters: target_id (required), new_title (required), new_content (required), update_reason. An unknown target_id is rejected.
• knowmind_upsert_entity: creates or updates a typed entity (person, company, product) and its aliases. Parameters: name (required), entity_class (required, see get_schema), description, aliases (string array). Classes outside the schema are rejected.
• knowmind_link_entities: creates a typed edge between two entries, with a confidence value. Parameters: from_id (required), to_id (required), rel_type (required, see get_schema), confidence (0–1 number). The reverse edge is created automatically.
• knowmind_unlink_entities: removes a typed edge. Parameters: from_id, to_id, rel_type (all required). The entries themselves remain.

There is no delete tool, and this is intentional. Facts are superseded via update_fact, edges are removed via unlink_entities, and history remains auditable. Deleting an entire workspace and its data is a self-service action in the knowmind.de web interface, not something an agent can do by calling a tool.

Old tool names On September 12, 2026, seven tools were renamed so thatEach name follows knowmind_<verb>_<object> and indicates read or write. Old names will continue to respond as-is until March 31, 2027, and will still appear in discovery, marked as deprecated. knowmind_entity becomes upsert rather than get: it creates a node or returns an existing node. Otherwise unchanged—same parameters, same behavior.

GET https://knowmind.de/api/mcp/v1 returns the name, version, protocol version, and the full tool list with input schemas without a token. Directory crawlers and inspectors can read the full interface anonymously; only tools/call requires a key.

Applicable scenarios

• Preserve project decisions, conventions, and facts across sessions in AI coding assistants (Claude Code, Cursor, etc.).
• Need to answer historical state queries such as "what did we believe in a certain year and month."
• Need a memory service with provenance, confidence, and auditable history, with data located in the EU.
• Store meeting minutes, specifications, and manual chapters as searchable documents.

Limitations

• knowmind is a hosted service, with no self-hosted community edition. If self-hosting is needed, the official recommendation is to consider other projects.
• The free tier includes 2,500 memories, 1 user, and 30 days of access logs.
• There is no delete tool; deleting an entire workspace requires self-service operation in the web interface.
• Content suspected of storing passwords or access keys will be rejected (422); pointers should be stored instead of keys.
• Specific pricing and current tier details are at knowmind.de/pricing; the materials do not provide a complete price list.

Documentation, pricing, data processing agreements, and privacy information are at knowmind.de. The German version is in README.de.md.