Builder open-sources agent memory that tracks abandoned decisions, not just facts
A common failure in AI agents isn't forgetting things — that's annoying but obvious. The harder problem is when an agent confidently re-proposes an approach that was already tried and abandoned a month earlier, or plans against a decision that was replaced two weeks ago because the old decision still looks just as valid in its notes. Nothing fails loudly; the agent just quietly wastes hours redoing work.
Generic s fix forgetting, but they don't fix being confidently wrong about the past, because that isn't a recall problem — it's a problem of tracking the current status of information (ruled out, replaced, still unverified). To address this, the builder spent three months creating NodeDex, a of a project's reasoning that gets built automatically in the background from the agent's conversations, so the agent never has to remember to save anything itself. The design treats dead ends as first-class data: an explicit, checkable list of approaches that were tried and abandoned, which the agent is taught to consult before proposing anything new.
Each decision is also stored along with its reasoning and the alternatives that were considered. The project has been released as .
Key points
- Agents confidently re-proposing abandoned decisions is a mode that causes repeated wasted work
- The core issue is tracking the current status of information (ruled out, replaced, unverified), not recall
- NodeDex is an tool that auto-builds a project reasoning graph from agent conversations in the background
- Abandoned approaches are stored as an explicit, checkable list the agent consults before proposing new ideas
- Each decision is stored with its rationale and the alternatives considered