New agent-memory design moves retained state into the behavior-selection layer
nice2Bnice2 · reddit · 2026-07-27
The team behind this system says it stopped treating agent memory as a storage/retrieval layer and instead moved it into the behaviour-selection path.
Architecture
- Input/current state
- Candidate behaviours
- Retained-state scoring
- Governor
- Final behaviour selection
The core idea is that memory storage and memory influence are different problems. The model can still generate the same candidate behaviours, while middleware applies controlled weighting to retained state before the final choice.
What it enables
- Compare Studio mode vs Governed mode on the same scenario and candidate set
- Keep the baseline model unchanged while changing final behaviour through retained-state weighting
- Remove the retained-state influence and fall back toward the baseline without retraining
- Preserve retained state across restarts
- Deterministic seed replay
- Explicit degraded/fallback reporting
- An inspectable end-to-end middleware path
The system is called Collapse Aware AI and the team says Phase-1 Core Gold Build is working, with integrations now being explored in agents, simulations and other stateful systems.
More from coding & agent
- Grok Build and Unity CLI/MCP push an AI game from scaffold to polished demo — Daniel_Farinax · 2026-07-27
- opencode v1.18.7 fixes macOS fullscreen inset and project picker scrolling — opencode-agent[bot] · 2026-07-27
- Andrew Ng memo argues agent memory needs graphs, not bigger context windows — solyarisoftware · 2026-07-27
- OpenSpace targets a skill management layer for AI agents, not just more skills — solyarisoftware · 2026-07-27
- Can Chrome’s installed LLM drive a customer chat and call MCP tools? — DashinTheFields · 2026-07-27
- An unused gaming PC becomes a remote Claude Code workspace — rchardkovacs · 2026-07-27