MemFlywheel: Giving Your AI Agents a File-Native Long-Term Memory

#ai #memory #agents #opensource

The Missing Piece in Today's Agent Hype

If you look at today's GitHub Trending, the conversation is dominated by execution (like browser-use automating video editing) and perception (like olmocr cleaning PDF data for RAG). These are critical, but they miss a fundamental component of intelligence: Memory.

An Agent that cannot remember its past interactions, learned skills, or context between runs is limited to single-turn tasks. To build truly autonomous workflows, agents need a "memory flywheel".

Introducing MemFlywheel

MemFlywheel is a file-native long-term memory layer designed for Agent Harnesses. It is not a replacement for your LLM or your tooling; it is the foundation component that enables your Agent to recall, extract, and evolve.

How It Works

MemFlywheel integrates directly into your Agent Harness (the host that owns lifecycle, models, and tools) to add three key capabilities:

Pre-recall: Before the agent acts, it scans MEMORY.md index cues to retrieve relevant past context.
Progressive Read: It layers reads from index cues to actual memory bodies and source traces.
Post-run Learning: After execution, it extracts durable memories and consolidates "dreams" (idle consolidation) to repair or refine memory.
Skill Flywheel: Repeated workflows are automatically evolved into reusable "learned skills" that the agent can inspect and reuse in future tasks.

Harness-Native & Open

MemFlywheel is designed to be agnostic to the host. It supports popular Agent Harnesses through npm packages:

Pi: pi install npm:@iflytekopensource/adapters
Hermes: memflywheel-hermes-install
OpenCode: opencode plugin @iflytekopensource/adapters --global
OpenClaw: openclaw plugins install npm:@iflytekopensource/adapters

By keeping memories as file-native Markdown, MemFlywheel ensures that memory is inspectable, diffable, and transparent. You can see exactly what the agent remembered and why, which is crucial for debugging and enterprise compliance.

Why This Matters Now

As the industry shifts from generic chatbots to vertical, execution-heavy Agents (as seen with the rise of RPA and specialized agents), the cost of statelessness becomes apparent. MemFlywheel allows developers to build Agents that:

Learn from experience: They get better at your specific workflows over time.
Maintain context: They remember preferences and constraints across sessions.
Operate securely: File-native storage allows for easy auditing and local deployment.