Mem Context

A vector-backed memory layer that persists context across AI assistant sessions using LanceDB and Ollama embeddings. Exposes six tools (recall, remember, forget, delete, purge, status) to store and retrieve conversation history, code diffs, error traces, and architectural decisions. Memories decay exponentially over time unless reinforced by access, with configurable rates per type (episodic fades fast at 0.15/day, semantic persists at 0.03/day, permanent never decays). Includes automatic deduplication via cosine similarity and a three-phase LLM-driven consolidation pipeline that extracts conclusions from sessions, merges them into permanent knowledge, and archives what goes unused. Ships with auto-provisioning for Claude Code and OpenCode, installing capture hooks that archive transcripts after each session without manual intervention.