MemFlywheel

What MemFlywheel does

MemFlywheel adds a file-native memory loop to agent harnesses: recall before execution, extract after execution, and evolve repeated workflows into learned skills. Memories are Markdown files with source traces, so they stay inspectable and diffable. The host harness keeps ownership of lifecycle, models, auth and tools, while MemFlywheel handles the memory and learning loop.

Recall is progressive: a pre-recall step reads MEMORY.md index cues, then memory bodies, source traces and learned skills as needed. At turn end it extracts durable memories, and while idle it runs dream consolidation and repair. Embedding pre-recall is optional. Without it, MemFlywheel injects up to 200 generated MEMORY.md index lines; with an OpenAI-compatible embeddings endpoint configured, it injects only the most relevant entries. One npm package supports all four hosts, and the Hermes route installs a MemoryProvider.

Key features

• Recall before execution, extraction after, and idle dream consolidation
• Markdown memories with source traces that stay inspectable and diffable
• Learned skills created from repeated workflows
• Optional embedding pre-recall through an OpenAI-compatible endpoint
• One npm package for Pi, Hermes, OpenCode and OpenClaw
• LoCoMo-oriented regression checks

When to use it

• Giving Hermes a memory provider that learns from each run
• Auditing agent memory as plain Markdown files in version control
• Sharing one memory approach across Hermes, OpenCode and OpenClaw

Who it is for: Agent builders who want auditable long-term memory that works across several agent harnesses, including Hermes.

How it fits with Hermes Agent

Supports Hermes Agent among several hosts, installing as a Hermes MemoryProvider selected with memory.provider.

How to install MemFlywheel

These commands are copied from the project's README. Check the repository for the latest steps before you run them.

npm install -g @iflytekopensource/memflywheel
memflywheel-hermes-install
hermes config set memory.provider memflywheel

Requirements: npm; optional OpenAI-compatible embeddings endpoint set through MEMFLYWHEEL_EMBEDDING_ENDPOINT, MEMFLYWHEEL_EMBEDDING_API_KEY and MEMFLYWHEEL_EMBEDDING_MODEL

FAQ