hippocampus-memory_skill

Overview

This skill implements a persistent memory system for AI agents that automatically captures, scores, reinforces, and decays memories—modeled on the hippocampus. It creates structured memory files, a weighted index, and scripts to encode, recall, and apply gradual decay. The system integrates with agent workflows to provide continuity across sessions and improve contextual answers over time.

How this skill works

The pipeline extracts signals from transcripts, assigns an importance score, runs a semantic check against existing memories, and either reinforces or creates concise summaries. Importance decays over time using a 0.99^days formula, and repeated mentions automatically bump importance and update lastAccessed. Scripts provide encoding, recall, daily decay, dashboard generation, and session core loading so agents can start with high-importance memories.

When to use it

• Add persistent continuity to conversational agents so they remember user preferences and past decisions
• Automatically capture important phrases like “remember that” and emotional or decision moments
• Load core memories at session start to provide immediate context
• Search memories with importance-weighted recall for context-rich responses
• Run periodic encoding and decay to maintain a relevant memory set

Best practices

• Use the provided cron jobs to run encoding every few hours and decay daily to keep importance scores current
• Store workspace under $WORKSPACE (default ~/.openclaw/workspace) so scripts and dashboard paths remain consistent
• Tune initial importance thresholds if your agent needs more or less aggressive core loading
• Keep HIPPOCAMPUS_CORE.md synced for RAG integration with OpenClaw memory search
• Review the Brain Dashboard periodically to spot low-importance archives or missing complementary brain skills

Example use cases

• Remembering user communication style and preferences to tailor responses across sessions
• Reinforcing facts about ongoing projects so the agent can surface relevant context without re-asking
• Archiving decisions and action items from meetings for later retrieval
• Loading core personal facts (identity, role, preferred format) at session start for consistent behavior
• Using recall.sh to pull background facts when answering context-sensitive queries