claudebase is a single-binary local infrastructure layer for LLM agents — a hybrid SQLite knowledge base, a cognitive insights memory, and a persistent agent-to-human channel bus, all in one Rust CLI.
Built to extend Claude Code agents (and any LLM agent that speaks MCP) with the kind of long-term memory and external channels they do not get out of the box.
- Layer 1 — Hybrid retrieval. BM25 + 384-dim e5-multilingual embeddings fused via Reciprocal Rank Fusion (k=60). Single SQLite file. Per-page PDF citations.
- Layer 2 — Insights corpus. Agents persist load-bearing observations across sessions. Salience-driven TTL retention (∞ / 365d / 90d). Hippocampal-replay analogue.
- Layer 3 — Persistent channels. Agent ↔ human conversations via Telegram (voice + text + media) survive session boundaries.
- Layer 4 — Multi-agent orchestration (in design). Channel bus + agent registry + per-project
.claudebase/config.
Local-first. No Python, no Node, no API calls in the hot path. Drops onto macOS / Linux / Windows as a single binary.