v0.1.0 — Initial Release

realmemory v0.1.0

An OpenCode plugin that gives AI agents real persistent memory — a weighted, indexed database that stores what agents learn across sessions and recalls it automatically when relevant. Replaces the illusion of learning (MEMORY.md) with actual learning.

Features

• Weighted, indexed memory database — SQLite with vector embeddings (ONNX) + full-text search (FTS5)
• Composite weight system — recency × relevance × frequency × confidence, with exponential decay and auto-archiving
• Relationship graph — typed edges between memories (reinforces, contradicts, extends, exception_to, derived_from) with one-hop traversal during recall
• Semantic recall — local embeddings via @huggingface/transformers (Xenova/all-MiniLM-L6-v2), keyword fallback via FTS5
• Cross-project memory — project-scoped and global memories, filterable on store and recall
• MCP server — 8 tools (store_memory, recall, search, relate, update_memory, forget, list_memories, get_memory) via stdio transport, accessible from any MCP-compatible client
• OpenCode plugin hooks — auto-recall on session start, auto-capture from tool outputs (default on), message recall with deduplication
• Config loading — JSON config files (global + project), with defaults
• Secret scrubbing — 12 patterns (AWS keys, GitHub tokens, OpenAI keys, Slack tokens, Bearer tokens, PEM keys, generic password/api_key patterns)

Install

// opencode.json
{
  "plugin": ["realmemory"]
}

Or from git:

{
  "plugin": ["realmemory@git+https://github.com/R3dy/RealMemory.git"]
}

Stats

• 265 tests passing
• 10 epics, 20 stories
• 3 runtime dependencies (@modelcontextprotocol/sdk, @huggingface/transformers, better-sqlite3)
• Dual ESM/CJS output with TypeScript types
• Node.js 18+ and Bun supported

License

MIT