Agent Continuity Curator

Maintain durable, lean, and consistent AI agent memory across sessions while preventing context bloat and data leaks.

Install
cmdop skills install agensi-agent-continuity-curator

Agent Continuity Curator is a skill that addresses the problem of short-lived session memory in AI agents. Without persistent memory management, context windows fill up with redundant chat history and stale project details, causing agents to lose behavioral consistency between sessions. This skill solves that by building and maintaining a structured set of durable memory files: SOUL.md for agent persona, USER files for user preferences, and WORKING_MEMORY.md for active project state.

The skill compresses long chat histories into compact, actionable summaries, pruning noise and enforcing strict memory budgets so token usage stays lean. It supports two operating modes: a Hermes-native mode that respects memory limits and delimiters, and a portable mode designed for high-capacity contexts such as Claude, Cursor, and local agent runtimes.

A built-in safety audit reviews memory content before persistence, excluding personally identifiable information, API tokens, secrets, and any unnecessary personal data. When an agent session ends, the skill generates a concise handoff summary that lets a new agent instance resume from exactly the point the previous one stopped.

This skill is appropriate when an agent needs to retain coding style decisions, architectural preferences, or user-specific context across multiple sessions without accumulating stale details. It is not a database or vector store, and it does not provide retrieval over large knowledge bases — its scope is structured working memory for a single agent’s continuity.

Use cases

  • Preserve an agent's coding style and architectural decisions across multiple work sessions
  • Compress a long chat history into a compact handoff summary when approaching context limits
  • Audit accumulated memory files to strip out PII, tokens, and secrets before persistence
  • Keep user preference profiles consistent across separate agent conversations
  • Enable a new agent instance to resume a project without re-reading full chat history
  • Maintain separate persona, user, and project-state memory files with enforced size budgets

When to use it

  • The agent runs across multiple disconnected sessions and must retain behavioral consistency
  • Context windows are filling with redundant or stale history that reduces response quality
  • Compliance or privacy requirements demand that PII and secrets are excluded from persisted memory
  • The deployment targets Claude, Cursor, Hermes, or a local agent runtime that needs structured memory files

When not to use it

  • The use case requires querying a large external knowledge base or vector store
  • A full database or long-term document retrieval system is needed rather than working-memory files
  • The agent runs in a single stateless session with no continuity requirement
  • The runtime does not support reading or writing local markdown-style memory files