Labsco
MCP SERVER

Frinus

by frinus-ai

Long-term memory for an agent, hosted: store episodic, semantic and procedural memories and retrieve them by task type.

Agent Memory
Summary
Your memories live on their servers, and the credential broker is the standout design choice.

This is a client for a hosted platform, not a local store — worth knowing before you write anything sensitive into it. Against that, `credential_exec` is genuinely well thought out: rather than handing a secret to the model, it injects the credential into a child process's environment, runs the command with no shell, and returns only the output. That is the right shape for the problem.

What it is

A bridge between your client and the Frinus platform, which is where the memories actually live. The server itself stores nothing locally — it validates your API key at startup, resolves your account and organisation, and forwards every call over HTTPS. What that buys you is a memory model with structure: three memory types, four visibility scopes, working memory with a size limit, sessions, a four-level knowledge hierarchy, and background consolidation.

What you get
  • 70 tools across memory, sessions, agents, knowledge hierarchy, orchestration and training
  • Memory: `memory_store`, `memory_search`, `memory_get`, `memory_list`, `memory_delete`, plus `memory_reinforce` and `memory_weaken` to raise or lower a memory's relevance after it helped or misled
  • Retrieval that knows what you are doing: `search_with_attention` weights memory types by `task_type` — debug leans episodic, documentation leans semantic, deploy leans procedural
  • Working memory: `working_memory_get`, `working_memory_add`, `working_memory_clear`, capped at seven items per context with oldest-first eviction
  • Sessions: `session_start`, `session_end`, `session_context`, `session_summary`, `session_clear`, with parent sessions for subagents
  • Continuous capture: `stream_capture`, `stream_get_session`, `stream_get_recent`, `stream_process` to promote important items into permanent memory
  • Housekeeping: `consolidation_detect_conflicts`, `consolidation_resolve_conflict`, `consolidation_detect_redundant`, `hierarchy_consolidate`, `sleep_run`
  • A knowledge hierarchy of concepts, themes, topics and points, each with create/list/update/delete
  • A credential broker: `credential_exec` runs a command with the stored secret injected into the child process environment and returns only stdout, stderr and the exit code, so the value never reaches the model
Requirements

A Frinus account and an API key shaped `sk-frinus-...`, generated under Settings → API Keys and shown once. Set it as `FRINUS_API_KEY` and leave the service URLs alone unless you self-host. Run it with `npx -y frinus-mcp@latest` — nothing to install. Package `frinus-mcp`, version 3.6.0. Without a valid key the server refuses to start. The free plan covers 100 memories and 20 queries a day, no card required.

Setup effort

One command plus a key — npx -y frinus-mcp@latest, then supply credentials