Changelog

Versioned changes to the Mnemexa public API, the official Python SDK, and the MCP adapter. Newest first.

2026-05-15 — Surface realignment

Major alignment pass: the four customer surfaces (backend endpoints, Python SDK, MCP adapter, docs) now mirror each other exactly.

mnemexa (Python SDK) — 0.1.1

Added

• client.status() and AsyncClient.status() → GET /v1/status. Returns workspace identity (id, name, current status, plan, key prefix) for connection / setup verification. Suspended and limit-reached workspaces still receive a 200 with workspace_status reflecting their real state.
• StatusResponse Pydantic model.

Removed

• client.events.ingest() and the EventIngestRequest / EventIngestResponse models. The underlying /v1/events/ingest endpoint is internal observability plumbing (single global relay URL, not per-workspace) and was never customer-facing. The 0.1.0 wrapper also had a wire-format mismatch with the backend that produced 400 on every call.

@mnemexa/mcp (MCP adapter) — 2.0.4

Fixed

• brain.health now reads the correct response fields from GET /v1/optimize/health. Bug present since 2.0.2 — the tool was reading top-level health_score / total_memories / stale_count, but the real response has quality_score at top level and counts nested under signals.*. Result was a “Memory Health Report” that always displayed 0% / 0 / 0 regardless of workspace state. Now reports quality_score on a 0–100 scale plus full signal breakdown (stale, duplicates, overlong, never-retrieved, open recommendations).

@mnemexa/mcp (MCP adapter) — 2.0.3

Changed

• brain.status now calls the new public GET /v1/status endpoint and returns the real workspace name, status, plan, billing cycle, and API-key prefix. Previously it returned a hardcoded “connected” message that didn’t verify anything against the backend. Non-active workspaces (suspended, limit_reached) now surface that state explicitly so users can self-diagnose.

Backend — public API

Added

• GET /v1/status — connection check returning workspace identity. Free diagnostic; no billing, no usage event, no lifecycle guard. Companion to /v1/optimize/health with the same semantics.

2026-05-15 — Initial public release

mnemexa (Python SDK) — 0.1.0

Initial public release. Sync and async clients wrapping four endpoints, Pydantic v2 typed responses, automatic retries for 502/503/504, error taxonomy, PEP 561 py.typed marker, Python 3.9–3.13 support.

@mnemexa/mcp (MCP adapter) — 2.0.2

Fixed

• brain.remember now sends { text } in the request body (was { content }), matching the backend POST /v1/memory/store contract. Previously every remember call returned 422.
• brain.health works against the newly-shipped public GET /v1/optimize/health endpoint (was 404 previously).

Removed

• brain.answer tool removed. The underlying /v1/memory/reason endpoint is out of scope. Use brain.recall to retrieve memories; synthesis is the calling model’s responsibility.

Backend — public API

Added

• GET /v1/optimize/health — public memory-health diagnostic. API-key auth, no billing, no usage event.

2026-05-10 — BizXEngine → Mnemexa rebrand

• Customer-facing brand changed to Mnemexa: mnemexa.com, app.mnemexa.com, api.mnemexa.com, docs.mnemexa.com.
• API key prefix changed from bxk_ws_ to mnx_ws_. Legacy bxk_ws_* keys continue to work — no rotation required.
• npm package renamed from @bizxengine/mcp to @mnemexa/mcp. Legacy package (@bizxengine/mcp@1.2.0) left undeprecated on npm as silent backward-compat; the legacy api.bizxengine.com host 301-redirects to api.mnemexa.com.
• Brand positioning: Mnemexa is now the Intelligent Memory OS for AI.

Internal system name remains BizXEngine for infrastructure, request headers (X-BizX-*), and database labels. See Authentication and Webhooks for header naming.