A refined but simplistic approach to agent memory

I work a ton across projects on my personal computer. I have tried Graphify, Graft and many others. None of them really fully solved the entire problem, but they did have parts. So, I built one that actually fit my needs. This removes the need entirely to choose or start sessions in specific repositories. It takes graph-based knowledge systems but strengthens retrieval, how it is updated, and much more. Effectively, it becomes a resilient system that you can actually watch your agents rely on, rather than something that just exists and is used only occasionally. It begins with the same base graph store, the exact same vector engine as Graft. I only built the orchestration layer on top of it, which makes it much more practical. Full attribution to them for this part. Instead of querying for a single-hit result, I found that it was much better for agents to have ranked retrieval and an actual graph walk. In practice this saves you chains of tons of bash greps and cds. It effectively gives the agent a trustworthy, probabilistic ranking of what is needed, with matching for strong, weak, stale, and rebuilt, based on lexical coverage and semantic matching. The code is not sloppy for this, it is personally edited. Both Graft and Graphify are solutions for single repositories or daemons . Heimdall is a layer on top it that watches the agent sessions, syncs the graph, and makes the retrieval trustworthy. Check it out at: github.com/ArihantDeva/heimdall MIT-licensed, with extensions: the verifier, the self-healing graph watcher, and the Graft adapter, with attribution. If you run agents across multiple projects, this is the missing layer.