elite-longterm-memory_skill

Overview

This skill is the ultimate long-term memory system for AI agents, combining WAL protocol, vector search, git-notes, curated archives, cloud backup, and auto-extraction. It preserves session state, captures decisions silently, and enables semantic recall so agents never lose context or repeat mistakes. Built for Cursor, Claude, ChatGPT and Copilot workflows.

How this skill works

The system writes a durable SESSION-STATE.md (hot RAM) before responses using a Write-Ahead Log to avoid lost context. Relevant facts are indexed into a vector store (LanceDB) for semantic recall, while structured decisions and lessons go to a git-backed notes graph (cold store). Human-readable distilled memory lives in MEMORY.md and daily logs, with optional SuperMemory cloud backup and Mem0 auto-extraction for fact deduplication and token savings.

When to use it

• When you need durable session continuity across restarts or compaction
• When agents must recall past decisions, preferences, or blockers reliably
• When you want semantic search over historical context for better responses
• When you need an audit trail of decisions stored silently and versioned
• When reducing token usage by auto-extracting facts is important

Best practices

• Always write to SESSION-STATE.md BEFORE responding (WAL protocol) to guarantee durability
• Capture important decisions silently into git-notes so outcomes are versioned and searchable
• Enable LanceDB auto-recall with conservative minScore to avoid irrelevant injections
• Keep MEMORY.md as a curated summary (<5KB) and link to detailed daily files
• Run weekly memory hygiene: audit vectors, archive daily logs, and clear low-value items

Example use cases

• Agent receives a user preference and persists it so all future sessions honor that preference
• Team decides a tech stack: record the decision in git-notes and auto-inject when relevant design discussions happen
• Long-running project: daily logs and MEMORY.md provide onboarding context for new contributors
• Multi-device workflows: SuperMemory backup syncs memory across agents and clients
• High-volume chat: Mem0 auto-extraction reduces token costs by storing distilled facts instead of raw transcripts