AgenticVMM: Distributed Memory Engine

Inspiration

Autonomous agents need to explore multiple paths, but traditional frameworks duplicate message history as strings. This ( O(n) ) bottleneck freezes systems and exhausts Edge VRAM. We built AgenticVMM to fix this by combining decision tree exploration with OS-level fork() mechanics to achieve instant, memory-efficient branching.

What it does

AgenticVMM is a hardware-level memory engine that eliminates re-prefill latency. By diving into the llama.cpp C-API, it physically clones KV-Cache pointers in the VRAM.

• ( O(1) ) Branching: New branches created in ~0.09ms.
• Zero VRAM Bloat: Newly created branches cost 0 MB additional VRAM.
• Persistent Backbone: Successful agent states are committed to CockroachDB via Managed MCP, with logs archived to AWS S3.

How we built it

We utilized a Python wrapper for llama.cpp's ctypes to manipulate VRAM at the pointer level.

1. CockroachDB MCP: For distributed state persistence and schema management.
2. ccloud CLI: Used to pin verified agent milestones into the cluster.
3. AWS S3: For long-term audit logs and state recovery.

Challenges we ran into

Handling C-pointers via Python caused frequent Segmentation Faults. Synchronizing high-speed VRAM operations with distributed SQL commits required a custom asynchronous fallback layer to prevent stalling the agent's decision loop.

Accomplishments that we're proud of

• 2450x faster branch creation compared to string-copying methods.
• Real-time persistence integration with CockroachDB.
• Memory-efficient operation on consumer hardware (6 GB VRAM).

What we learned

We learned the massive efficiency gap between logical text-based memory and physical matrix-based memory, and how vital Managed MCP is for maintaining agent "long-term memory" across distributed sessions.

What's next for AgenticVMM: Distributed Memory Engine

• Multi-Agent KV-Cache sharing.
• Semantic branch retrieval using CockroachDB vector embeddings.

Built With

• amazon-web-services
• c-api
• ccloud
• cockroachdb
• edge-ai
• llama.cpp
• llm
• mcp
• python
• s3