Nothing to wire up
Facts are extracted from your turns and injected on the next one. Same endpoint, same key, every public lane. No second client, no per-editor setup.
Memory is already
in the gateway.
Strix memory ships in production on the same strix_ key you already call: facts extracted and injected silently, scoped per project, with pin, recall, history, feedback, batch, MCP, and forget. Mem0 is the sidecar you would otherwise host next to it.
01 / WHAT STRIX SHIPS
Full CRUD + recall
Pin, list, get, patch, delete, search. Batch pin up to 20. Metadata and categories on every fact. MCP tools on the same Worker if you want them.
Auditable and forgettable
Per-fact history, thumbs up/down feedback, a sortable dashboard table across all projects, and GDPR forget by project or account.
| Strix ships | How | Mem0 equivalent |
|---|---|---|
| Silent extract + inject | Every public lane on your strix_ key — no client changes | SDK call per turn |
| Project scopes | x-strix-project or inferred git/workspace | run_id convention |
| Add / search / get / update / delete | /v1/memory* plus MCP tools | add / search / get / update / delete |
| Recall ranking | Vector + lexical + Cohere rerank | Vector search |
| Per-fact history | GET /v1/memory/:id/history add / update / delete / feedback trail | history(memory_id) |
| Feedback | POST /v1/memory/:id/feedback helpful or not, per fact | feedback |
| Batch pin | POST /v1/memory/batch up to 20 verbatim facts | batch add |
| Metadata + categories | ?category= filter, plus flat metadata set at pin time | metadata + categories |
| Custom extraction rules | extract_instructions on memory settings | custom instructions |
| MCP server | Same Worker, same key — nothing new to host | Separate hosted MCP |
| Session dashboard | Sortable table, learned vs pinned, all projects | Hosted platform UI |
| GDPR forget + export | Per project or whole account, tombstone + vector ids | delete_all |
02 / MEM0-ONLY, BY CHOICE
We do not clone these.
Strix memory is short derived facts for coding agents, not a retrieval platform. If you need any of the following, keep Mem0 — it runs happily alongside Strix.
| Mem0-only | Why we skip it |
|---|---|
| Entity / knowledge graph | Use Mem0 or Graphiti when you need relations |
| Notion / Drive / file connectors | Strix stores short facts, not a corpus |
| Document RAG | Bring your own retrieval layer |
| Python / JS memory SDK | The gateway is the API — HTTP or MCP |
| user_id / agent_id / run_id filters | Tenant is the API key; the extra axis is project |
Every query is scoped to the authenticated key. Event payloads store role only. Vectorize holds ids and tenant filters, not fact text. Memory docs · Compare index
FAQ
Straight answers
before you start.
Trial needs an OpenRouter key. Tokens stay on their bill. Strix is the router.
- Do I bring my own OpenRouter key?
- Yes. The 7-day trial and paid plans require an OpenRouter key (BYOK). OpenRouter bills model tokens at their rates. Strix sells routing, memory, and the control plane — not a token markup on BYOK inference.
- What am I paying Strix for?
- Nine public intent models, payload filtering, failover, a memory layer (cross-session facts, Cohere rerank, forget/export), and Team seats/keys/spend controls. Auto plus cheap lanes such as quick and coding keep everyday turns off always-Sonnet pricing.
- Which clients work?
- Any OpenAI-compatible client, plus Anthropic /v1/messages. That includes Cursor, Claude Code, Codex, OpenCode, Cline, and Continue — not Cursor only. Point the base URL at api.strixgate.dev and use a strix_* key.
- Is uncensored or Venice required?
- No. Uncensored is an optional lane. Venice is used only if you store a Venice key. Default routing stays on OpenRouter with your key.
- Do you log prompts?
- No prompt logging for training or human review of your request bodies. We meter usage, lanes, and spend. You can forget or export stored memory facts from the dashboard.
- How is this different from OpenRouter or Cursor alone?
- OpenRouter still bills tokens — you bring that key. Strix sits in front with lanes, filtering, and memory. Cursor is one supported client, not the only one. Use Compare in the nav for OpenRouter, Cursor, and alternatives pages.
- How does Team pricing work?
- Team is $99/month with three seats included, then $20 per extra seat. Builder is $29/month for one person. Both include a 7-day trial.