Agentix LabsInternal product architecture
Building shared agent memory without making every old message true
NeuraKeep was designed to give Claude, Codex, OpenClaw, and compatible clients cross-session context while preserving provenance, sensitivity, correction, and human review.
Disclosure: NeuraKeep is an Agentix Labs product. This case study documents the internal architecture and operating lessons.
The risk
Transcript memory mixes source facts, model guesses, secrets, obsolete instructions, and later corrections. If every item becomes durable automatically, the memory system amplifies errors instead of reducing them.
The architecture
Raw sourcesOwned files, notes, sessions, and artifacts remain reconstructable.
Cited retrievalDurable claims preserve source and cited passage.
Proposal reviewAgents propose facts, decisions, events, and failures; policy or a reviewer applies them.
Failure memoryKnown bad paths are searchable before another agent repeats them.
Implementation lessons
- Start with read-only MCP tools and require citations.
- Separate tenant, project, sensitivity, trust, and time scopes.
- Keep conflicts and corrections visible instead of overwriting history.
- Make export and rebuild possible from owned source material.
Evaluate your memory layer
Use NeuraKeep’s checklist before connecting autonomous writes or sharing memory across agents.
Open the evaluation checklist