Token Savior Recall

Token Savior Recall is an open-source MCP server that dramatically improves Claude's coding performance by replacing whole-file reads with symbol-level navigation and adding a persistent memory engine. It indexes codebases by symbol — functions, classes, imports, and call graphs — so the model navigates by pointer instead of reading entire files, reducing injected characters by up to 97%. A SQLite-backed memory engine stores decisions, bugfixes, conventions, and session rollups with Bayesian validity scoring, contradiction detection, and per-type TTL decay, re-injecting a compact delta at the start of each new session.

• Structural code navigation — Index your codebase by symbol and navigate via 90 tools covering core navigation, dependency graphs, git diffs, safe editing, checkpoints, and more; install via pip install "token-savior-recall[mcp]" or uvx token-savior-recall.
• Persistent memory engine — 21 memory tools backed by SQLite WAL + FTS5 + optional vector embeddings (all-MiniLM-L6-v2, 384d), with hybrid BM25 + vector search fused via RRF for cross-session knowledge retention.
• Progressive disclosure contract — Three-layer memory search (memory_index → memory_search → memory_get) keeps lookup cost bounded at ~15, ~60, and ~200 tokens per result respectively.
• Bayesian validity & contradiction detection — Each observation carries a validity prior with update rules; contradictions are flagged at save time and stale observations surface explicitly rather than being silently trusted.
• Configurable tool profiles — Six profiles (full, core, nav, lean, ultra, tiny) filter the advertised tools/list payload to match manifest budget constraints, from ~1,070 tokens (tiny) to ~8,770 tokens (full).
• Defer-loading via ts_search — The tiny profile exposes 6 tools and makes ~60 others reachable just-in-time via semantic tool search using Nomic 768d embeddings, mirroring the Tool Attention pattern for −95% prefix tokens.
• Benchmark-validated performance — Achieves 192/192 (100%) on the tsbench benchmark with Claude Opus 4.7, versus 141/180 (78.3%) for plain Claude, with zero losses across 96 tasks and 12 coding categories.
• Broad MCP client compatibility — Works with Claude Code, Cursor, Codex CLI, Cline, Continue, Windsurf, Aider, Gemini CLI, Copilot CLI, Zed, and any custom MCP client.
• Optional integrations — Telegram bot feed for critical observations, opt-in web viewer via htmx + SSE, and LLM auto-extraction (TS_AUTO_EXTRACT=1) for passive observation capture from tool use.