Contextual Memory Virtualisation

Contextual Memory Virtualisation (CMV)

Like virtual memory lets an OS use more RAM than it physically has, CMV lets an AI coding agent use more context than fits in one session — by snapshotting, branching, and trimming the understanding your sessions build up. The approach is agent-agnostic; this implementation demonstrates it with agentic coding sessions in Claude Code. For the full design rationale and evaluation, see the paper.

The Problem

Every conversation with Claude builds up state. Not just chat history — actual understanding. Architecture mapped, tradeoffs weighed, decisions made, conventions learned. After 30 minutes of deep work (sometimes less!), Claude has a mental model of your whole codebase sitting in its context window. That took real time and real tokens to build.

And then it's just... gone. The session context fills up, you run /compact, and the model condenses itself into a few sentences. Next session, you start from scratch.

Contextual Memory Virtualisation (CMV) treats that accumulated understanding as something worth keeping around. Named snapshots you can return to, branch from, and trim down — basically version control for context.

- Snapshot captures context state at a point in time (think git commit)
- Branch forks into a new independent session from that state (think git checkout -b)
- Trim strips dead weight from context without losing any of the actual conversation (no git equivalent — this one's new)
- Tree shows how your context evolved over time (think git log --graph)

Build up context once, reuse it whenever you need it.

What's actually happening inside:

Claude's synthesis — the part with actual understanding — stays. The raw file dumps, tool outputs, thinking signatures, and image blocks are gone. If Claude needs a file…