Idea: optional local-first persistent agent memory backend for RAI agents

[samvallad33]

Disclosure first: I maintain Vestige, so this is an interested-party idea, and per CONTRIBUTING I am raising it for discussion rather than opening a PR.

RAI agents accumulate context within a run, but persistent memory that survives across sessions, and that can decay or be corrected over time, is a harder problem on a robot where you also want it on-device. I wanted to float an optional local-first memory backend as something to discuss, not build unprompted.

Vestige - Local-first cognitive memory MCP server for AI coding agents. FSRS-6 retention, prediction-error gating, active forgetting, spreading activation, 3D dashboard. Single Rust binary. npm install -g vestige-mcp-server.

The relevant properties for an embodied agent: it runs fully on-device, decays stale observations on a real forgetting curve so memory does not grow without bound, avoids re-storing near-duplicate observations, and can hold a contradiction instead of silently overwriting. It speaks MCP, so if RAI's agent layer can consume an MCP tool it integrates without core changes.

I am explicitly not proposing to build anything yet. If the direction is interesting I am happy to scope an integration shape here first. If memory is already on your roadmap in a different form, equally happy to defer.

Sam Valladares, github samvallad33