hipocampus-flush

The hipocampus-flush skill provides a mechanism for AI agents to manually persist session data before reaching standard automatic checkpoints. In complex, long-running agent workflows, maintaining context is critical for preventing information loss and ensuring the agent's knowledge base remains current. By invoking this skill, the agent summarizes the active session—including user requests, analytical findings, decisions made, outcomes achieved, and relevant file references—and dispatches a subagent to commit this information to the daily memory log (memory/YYYY-MM-DD.md).

Following the write operation, the subagent triggers a compaction process. This is essential for the Hipocampus memory architecture, which relies on a multi-tier structure comprising hot storage, raw logs, and a hierarchical compaction tree. Compaction ensures that session insights are propagated through the index, reindexed by the qmd system, and made available for future cross-session retrieval. This tool is intended for developers and power users working with agent frameworks like Claude Code or OpenClaw who need to ensure high-fidelity context retention across extended development sessions.

• Performs a structured extraction of session state into a standardized markdown format.

• Automates the dispatch of a subagent to perform file I/O, preventing contamination of the active workspace and context window.

• Triggers tree propagation and qmd reindexing to maintain search performance and implicit context relevance.

• Ensures that crucial decisions and research findings are preserved in permanent storage, mitigating issues related to context window degradation.

• Invoke using /hipocampus:flush at any point during a session when significant progress has been made.

• Follow up with hipocampus:compaction if a full, immediate tree re-sync is required, though the flush skill often handles this internally via the subagent workflow.

• Ensure the environment is properly initialized with npx hipocampus init to allow the subagent to access the correct memory directories.

• This skill is highly effective for transitioning between unrelated tasks to prevent cross-contamination of working memory.