Cortex
Organizational Memory for AI Agents
Problem
Enterprise knowledge is fragmented across Slack, GitHub, and Jira. Agents that only see one silo invent context, lose provenance, and cannot be trusted across sessions.
Approach
Built a Python/FastAPI core with a Kafka extraction pipeline that pulls structured decisions into Neo4j via GPT-4o function calling.
Exposed memory natively to MCP-compatible agents through a TypeScript MCP server, with trust scoring, contradiction detection, and RBAC.
Tuned retrieval for multi-agent use across 10K+ nodes so context lands before inference, not after a failed answer.
Results
10K+ nodes. 30% recall lift. MCP native
- Cut multi-agent retrieval latency by 120ms across 10K+ nodes
- Improved cross-session context recall 30% over baseline prompting
- Trust scoring, contradiction detection, and RBAC on agent memory
Stack
Learnings
Agent memory is an infrastructure problem before it is a prompting problem. Provenance and access control decide whether teams trust what the agent recalls.