agent-memory

Persistent OKF-native hybrid-search memory for AI coding agents. Durable SQLite memories are canonical structured concepts; lossless Markdown documents, indexes, logs, and bundles are virtual handler projections. This provides selective retrieval, knowledge-graph traversal, cross-project search, and agent-scoped memories without turning a file tree or Git into the source of truth.

Architecture

• SQLite -- single-file backing store, portable, zero config
• OKF-native concepts -- every durable SQLite memory is canonical structured knowledge; Markdown is a lossless virtual projection, not the source of truth
• fastembed-rs -- local embeddings via all-MiniLM-L6-v2 ONNX model (semantic similarity, no API calls)
• Hybrid ranking -- BM25 (FTS5) + cosine similarity fused via Reciprocal Rank Fusion (RRF), then re-ordered by a local cross-encoder reranker (fastembed, on by default)
• MCP server -- stdio JSON-RPC server for native Claude Code tool integration
• CLI -- direct command-line interface for humans, scripts, and AI agents

OKF-native memory

Every durable memory is an OKF concept stored canonically in SQLite. Its body is ordinary Markdown-compatible text; normalized metadata, sources, verification, lifecycle state, typed relationships, extensions, and immutable revisions live in relational tables. memory okf handlers make that state act like Markdown files and bundles without requiring physical files or Git.