Changelog

knowmind CLI 0.1.25: guided onboarding. The new command knowmind init --token <kmt_…> --api https://knowmind.de wires up token and endpoint in a single step and prints a hint about completing the token setup at the end (knowmind login --token remains a valid alternative). It also adds a near-dup guard on save: paraphrased duplicates are detected before they are created — instead of producing a second entry, it points to the existing one.

GTM cut: API and CLI access is now available on every plan, including the free Private tier (previously only from Business API). Tokens are issued across all tiers; the memory limit on the Private plan is 2,500 memories. From now on, Business API differentiates itself solely through higher throughput (300 requests per minute), webhooks and priority — no longer through API access itself.

knowmind CLI 0.1.23: new command knowmind install <ide>, which wires knowmind into eight IDEs. The integration writes the matching MCP configuration per IDE schema, is idempotent and never stores secrets in plain text in the project.

Free tier: 2,500 memories + a fair soft limit. The Private plan now stores up to 2,500 memories free of charge (previously 100). The limit is now actually enforced: the count covers the stored documents/memories of the workspace — not the auto-extracted graph nodes. From 80% usage, knowmind_stats, store_memory and upload_document return the fields memoryLimit, memoriesUsed and warnAt80 plus a plain-text hint; the dashboard shows the usage and a banner from 80%. Once the limit is reached, the write paths reject with 402 plan_upgrade_required (MCP: a JSON-RPC error with the same message); recall, export and all existing content remain fully available.

Eleventh MCP tool knowmind_list_recent (latest corpus entries, created_at descending). Write path fixed: knowmind_store_memory and knowmind_upload_document now actually persist — Postgres document, vector index and graph node instead of a volatile in-process store; stored entries are findable via recall and list_recent and survive restarts. Documents: upsert-by-title (a new version replaces old versions only after a successful ingest). Errors now return a JSON-RPC error instead of a faked success. Tool reference, OpenAPI description and all integration pages updated to eleven tools.

Stability and consistency release after a full audit: MCP tool names unified to the knowmind_ (underscore) form across the remote endpoint, the npm package and the docs; streamable-HTTP responses are now spec-compliant (standard message events), fixing hanging tool calls in Claude. OAuth connectors are multi-tenant, with scope validation and rate limiting on the token endpoint. Magic links are valid for 24 hours. Transactional mail fully moved to our own SMTP (ALL-INKL.COM, Germany). Stripe webhooks extended (plan changes, failed payments), prices shown as final prices incl. VAT. Timeouts on all backend calls.

Full restructure along the Diátaxis model (tutorial, how-to, reference, explanation). Security section added to ISO/IEC/IEEE 82079-1:2019. New integration guides for Claude Desktop, Claude.ai, ChatGPT and Cursor with OAuth and token options. Complete MCP tool reference with ten documented tools. Glossary expanded past 40 entries. English mirror released under /docs/en.

Private plan capped at 100 memories. Forum bot renamed "Documentation Search". Audit retention now tiered by plan: 30 days (Private) up to 5 years (Enterprise). Rate-limit headers hardened, wikilink extractor added to the ingestion path.

knowmind CLI 0.1.5: preference / fact / update extractor before ingestion, which stores tagged additional chunks. Sync command with manifest and content hash. Package published on npm.

knowmind established as the brand. Control plane live at knowmind.de with magic-link sign-in via Brevo. Docs moved to /docs, Diátaxis sections introduced, OAuth for custom connectors enabled.

Sprint 1 gate exceeded: Recall@5 reaches 100 percent on the internal gold standard, P99 latency 239 ms. Corpus migrated to the sc-internal tenant; multilingual-e5-large becomes the default embedding model.

Hybrid recall with BM25 (German analyzer), pgvector and Neo4j graph hops + cross-encoder rerank promoted to the default path. Recall quality test reaches 98.61 percent on the internal corpus.