| name | enterprise-agent-memory-skill |
| description | Build cross-session persistent memory for enterprise agents that auto-captures tool usage, generates semantic summaries, injects relevant context in future sessions, supports tiered retrieval with token cost visibility, skill-based search, privacy controls with sensitive content exclusion, and audit trails. Use when Codex must design, provision, or generate an enterprise agent memory system with hook-based capture, AI-powered compression, ChromaDB or CSS/OpenSearch vector retrieval, progressive disclosure, 5-layer memory architecture (L1 working, L2 episodic, L3 skill tree, L4 semantic index, L5 config), FTS5 SQLite retrieval, Markdown memory files, self-improving memory loops, Huawei Cloud deployment, or privacy-compliant agent memory. |
Enterprise Agent Memory
Use this skill to produce implementation-ready agent memory systems, not generic chatbot advice. Keep outputs privacy-compliant, auditable, and token-efficient.
Overview
Enterprise agents lose context between sessions, forcing users to re-explain background every conversation. This skill provides cross-session persistent memory that auto-captures tool usage, AI-powered semantic compression, and injects relevant context in future sessions.
Inspired by:
- claude-mem — Hook-based capture + AI compression (~400 input + 150 output tokens/call) + ChromaDB vector retrieval + progressive disclosure
- hermes-agent — FTS5 SQLite retrieval + Markdown memory files + self-improving loop (memory + skills + soul + crons)
- GenericAgent — 5-layer memory (L1 working, L2 episodic, L3 skill tree, L4 semantic index, L5 config) + self-evolving skill tree
Core value: every session starts with relevant context already loaded. Token consumption drives the business model — each agent execution produces capture, compression, retrieval, and injection costs.
Default Workflow
- Install capture hooks on agent tool calls (lifecycle events: tool_start, tool_end, session_start, session_end, error).
