Three products, one engine
Run the memory framework yourself, let us host your agents, or let us run the memory layer entirely. Every path uses the same open-source build.
NMAFC Memory Framework
An open-source memory layer for LLM agents. It compacts each turn into dated atomic facts, ages unused ones on a decay curve set for the long haul, and retrieves through a fusion of five ranked channels — so context stays small no matter how long the conversation gets.
- Temporal grounding: relative dates resolved at write time
- Three memory types with different decay rates
- Source attribution back to the originating turns
- No external vector database required
Agent Hosting
Pick a harness and we provision it, run it, give it a subdomain and wire it to the same memory. OpenClaw and Hermes Agent and OpenCode are available now; Claude Code, Codex, Gemini CLI are announced.
- OpenClaw, Hermes Agent, OpenCode: available now
- Claude Code, Codex, Gemini CLI: announced
- Desktop editors connect over MCP instead
- Swap harnesses without losing what was learned
Hosted Memory Service
The same engine, operated by us. Multi-tenant and isolated by credential on our infrastructure, or deployed into your own VPC, on-premise, or air-gapped. You bring an API key; we bring the operations.
- Managed multi-tenant: built and tested, not yet deployed
- Your VPC, on-premise and air-gapped: announced
- Vectors stay in-region, but turns still reach your LLM provider
- Enterprise controls published with current state
Metered LLM Access
Use our model keys with per-request metering, or bring your own. The same proxy fronts every major provider, so switching costs nothing. Platform credits are billed against a credit balance that floors at zero rather than going negative.
- Real-time token metering
- Credit balance with a hard floor
- Bring-your-own-key supported
Start with the framework
It is free and open source. Host an agent or the memory layer only when you want us to operate it.