Agent memory worked better as a statement graph than as a bigger context window
chrislally · reddit · 2026-07-26
The author found that agent memory kept failing when they tried to make it work by stuffing more context into the window or searching harder over history.
They switched to a statement graph model with typed entities, relationships, and statements about statements, so each fact can carry provenance, validity windows, and supersession history. The post lays out several lessons:
- framework memory assumptions break under real-world identity conflicts and temporal validity
- names should be treated as labels, not IDs
- temporal history should be appended, not overwritten
- conflicting claims should remain visible until a policy or human resolves them
- resolution depends on trust rules, so the believed result should be recorded too
The remaining open problem is context assembly: how to pick the right bounded slice of the graph during execution.
Related event: Andrew Ng Advocates Graph-Based Memory for AI Agents(2 posts)→
More from coding & agent
- Kolab Benchmark Uses Pro Screen Recordings to Show AI Agents Overestimate Collaboration Quality — minnesotanlp · 2026-10-27
- Dev argues Claude burns far more tokens than GPT models for the same coding work — haider1 · 2026-10-27
- Stop building agents that can't change models: stable contracts, external state, real evals — rseroter · 2026-10-27
- Cheap models can cost more in agent harnesses than frontier ones, researcher warns — vasuman · 2026-10-27
- Kanzo: A 5k-Line Rust Kanban Binary With MCP for Claude Code and Codex — allmudi · 2026-10-27
- 'A human approves everything' is a trap: approval fatigue and how to actually design safe agents — Efistoffeles · 2026-10-27