Elite Longterm Memory

• You lose context mid-session — the WAL (Write-Ahead Log) protocol ensures state is written before the agent responds, so compactions and crashes can't erase what matters.
• You work across long-running or multi-session projects — the curated MEMORY.md + daily logs give both agent and human a durable, human-readable record of decisions.
• Your agents repeat mistakes or forget preferences — Mem0 auto-extracts facts from conversations and deduplicates them, reducing token usage by ~80% vs. raw history.
• You run sub-agents or parallel workstreams — the layered architecture lets you pass structured context via git-notes and semantic search rather than re-explaining everything.

• You need a one-off, single-session task — the full 5-layer setup adds overhead that isn't justified for throwaway scripts or short tasks.
• You're not using an OpenAI-compatible LLM key — semantic vector search (LanceDB) and Mem0 integration require OPENAI_API_KEY; the skill won't reach its full potential without it.

State is written to SESSION-STATE.md before the agent responds, not after. This prevents context loss during compaction, crashes, or mid-conversation interruptions — a guarantee that naive memory approaches can't provide.

All memories are indexed in a local LanceDB vector store. Auto-recall injects the most relevant prior context into each response, and manual memory_store / memory_recall commands let you query or filter by category and importance score.

Structured decisions, learnings, and branch-aware context are stored as git notes — silent, version-controlled, and exportable as JSON. They persist permanently alongside your repo history.

Mem0 automatically extracts facts, preferences, and decisions from raw conversation history, deduplicates them, and surfaces relevant memories on demand. This eliminates the need to pass full chat history and cuts token consumption by up to 80%.

Optionally sync your knowledge base to SuperMemory for cross-device access and natural-language querying of stored context. Requires a SUPERMEMORY_API_KEY.

A MEMORY.md root file and dated daily logs under memory/YYYY-MM-DD.md give both agent and developer a clear, browsable record of what was decided, learned, and remembered — no opaque database required.

As the project evolves across days or weeks, the agent writes tech decisions (e.g. "Use Tailwind, not vanilla CSS") to git-notes and SESSION-STATE.md. On every new session it reads prior context automatically, so it never asks what framework you're using again.

When the context window fills and the agent compacts, SESSION-STATE.md acts as hot RAM that survives the reset. The WAL protocol ensures that anything said in the last exchange is already persisted before the compaction fires.

Context stored in git-notes and LanceDB can be queried by any spawned sub-agent at task start, so parallel agents share the same decisions and preferences without manual re-briefing.

User preferences, recurring deadlines, and corrected mistakes are auto-extracted by Mem0 and stored across sessions. The agent builds an accurate, deduplicated model of the user over time without growing the prompt.