Beyond Governance: Adding Persistent Memory to AI Agents with MemFlywheel

#ai #agents #memory #opensource

The Missing Link in Agent Governance

As the GitHub Trending today highlights, the industry is shifting focus from "skill explosion" to centralized governance, version control, and RBAC (like claude-skills and SkillHub). This is crucial for enterprise security.

However, there is another critical pain point: Context Amnesia.

Even with perfect permissions, if an Agent doesn't remember previous interactions, it cannot build on past work. It starts from zero every time. This is where MemFlywheel comes in.

What is MemFlywheel?

MemFlywheel is a file-native long-term memory layer for AI Agents. It is not just a chat history; it is a structured memory foundation component inside an Agent Harness.

Key Features

File-Native Memory: Memories are stored as Markdown files, making them inspectable, diffable, and human-readable. No black-box vector databases required for simple recall.
Progressive Recall: Before execution, it recalls relevant context from index cues. After execution, it extracts new information.
Skill Flywheel: Repeated workflows are automatically evolved into "learned skills," which can be reused or published to a registry like SkillHub.
Harness-Native: It integrates seamlessly with popular Agent Harnesses like Pi, Hermes, OpenCode, and OpenClaw via npm packages.

How It Works

The architecture follows a clear loop:

Pre-recall: Fetches memory index cues before the agent acts.
Post-run Learning: Extracts durable memory after the turn ends.
Dream Consolidation: During idle time, it consolidates memories and repairs inconsistencies.
Skill Evolution: Identifies repeated work and converts it into reusable skills.

Agent Harness
   |
   |  lifecycle / model / auth / tools
   v
MemFlywheel
   |
   |-- pre-recall       -> MEMORY.md index cues
   |-- progressive read -> memory bodies -> source traces -> learned skills
   |-- turn-end         -> durable memory extraction
   |-- idle             -> dream consolidation and repair
   `-- repeated work    -> reusable learned skills

Integration Example

For Hermes users, installation is straightforward:

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

Why It Matters for Governance

While tools like SkillHub handle who can use what skill (Governance), MemFlywheel handles what the agent knows (Memory). Together, they form a complete enterprise-grade Agent ecosystem:

SkillHub: Version control, RBAC, Audit logs.
MemFlywheel: Long-term context, continuous learning, skill evolution.

By combining governance with memory, we ensure agents are not only safe and compliant but also intelligent and persistent.