- 01activity
- 02entity
- 03suggestion
- 04human confirms
What it is
Provider-agnostic passive memory for agent systems — notices entities in activity streams, surfaces noteworthy suggestions, learns from dismissals, and deduplicates. Convex is one proof, not the only one.
Topics
- agent-memory
- agents
- convex
- llm
- typescript
From the README
Live graph rail: NodeMem notices entities as dim, unmeasured nodes with no edges; a human confirms one suggestion and exactly one faint traversal edge appears. Animated demo hidden because your device asks for reduced motion. Open the animation.
NodeMem/assets/graph-rail/live-graph-rail.gif · 960×590, 2.3 MB · in the READMEMore in agent runtime and memory
What an agent reads, remembers and records.
-
Agent runtime
Node
Agent The room asks. The agent answers — with sources, a model, and a memo.
- room context
- document lookup
- versioned sheet
- notebook memo
-
Semantic graph
Node
Graph Artifacts, traces and proposals into an evidence-backed graph.
- room data
- semantic graph
- select cluster
- render
-
Agent observability
Node
Trace Open the trace behind each tagged surface in an agent-built app.
- tag surface
- record trace
- trace lens
- proof card
-
Agentic RL
Node
RL Turn failed agent runs into the next better attempt.
- act
- evaluate
- remember
- repair
- export