Memmy

MemTensor's local-first memory hub and agent, on your key. Reaches: the whole catalog — it speaks both wires, so the same key serves open-weight chat ids and Claude ids depending on which provider you configure.

Point ~/.memmy/config.yaml at the gateway

Memmy resolves provider: against its own provider registry. It has no entry for InferenceHub, so use custom — the escape hatch for gateways — and override the endpoint with apiBase. The base URL ends in /v1:

# ~/.memmy/config.yaml
agents:
  defaults:
    model: zai-glm-5.2
    provider: custom
providers:
  custom:
    apiKey: sk-prov-live-YOUR_KEY
    apiBase: https://app.inferencehub.tech/v1

Memmy expands ${ENV_NAME} in this file, so you can write apiKey: ${IH_KEY} and keep the key out of the config entirely.

Run it

memmy agent -m "summarise what I worked on yesterday"

Tool calls route through the gateway unchanged, which is what the memory Skills are built on — so memmy-memory init and the per-agent adapters work the same way they do on a first-party provider.

Optional — use the Messages wire instead

Memmy's anthropic backend honours apiBase too. Point it at the bare origin (no /v1 — Memmy appends the Messages path itself):

# ~/.memmy/config.yaml — Messages wire instead
agents:
  defaults:
    model: glm-5.2
    provider: anthropic
providers:
  anthropic:
    apiKey: sk-prov-live-YOUR_KEY
    apiBase: https://app.inferencehub.tech

Troubleshooting

If you see…	Fix
`No API key configured for provider 'null'`	The name in `provider:` isn't in Memmy's registry — a typo, or a gateway name like `inferencehub`, which it doesn't know. Use `custom` (or `anthropic`) and set `apiBase`. The message names `'null'` rather than the bad value; that's reported upstream.
"Using default configuration." then a crash	The config failed to load and Memmy fell back to defaults instead of stopping, so the real error is the one printed first. Check the YAML — an unset variable referenced as `${ENV_NAME}` will do it.
"model not found"	Match the wire: open-weight chat ids (`zai-glm-5.2`, `gpt-oss-120b`, …) on `custom`, Claude and GLM ids on `anthropic`. Live ids are on the Models page.
`apiType is only supported for providers.openai`	Memmy only lets the built-in `openai` provider pin `chatCompletions` vs `responses`. Drop the line — `auto` resolves to Chat Completions for `custom`, which is what the gateway serves here.