Callimachus Recall
One local, searchable index of every AI coding-agent conversation you've had — Claude Code, Codex, Cursor, Gemini and more — so you never redo solved work.
The skill reads/writes only Callimachus's own index and memory, explicitly never edits user files or runs commands in user projects; SECURITY.md discloses the local-first data flow (keys in OS keychain, egress only to a user-chosen provider), and the four write tools are tightly scoped. Deducted for: writes to AGENTS.md/CLAUDE.md (cal agents, SessionStart hook) and recorded decisions persistently modify repo files without explicit user confirmation; conversation history is sensitive data — read scope is disclosed but isolation/rollback are not.
SKILL.md is internally consistent: cal CLI fallback when MCP is unavailable, explicit distinction between 'empty = not distilled' vs 'no history', and documented degradation to search_threads/get_thread — diagnosable failure feedback. Deducted for: static review cannot execute key paths; parameter/behavior consistency of 21 MCP tools is documentation-only; no skill-specific tests or error-code documentation.
Trigger conditions are precise ('did before', 'what did we decide', load project memory at task start) and boundaries are clear (own index only, semantic recall requires Ollama/cloud key, Cursor threads lack project paths). Deducted for: English-only documentation; first-run embedding-model download from Hugging Face and reliance on Ollama Cloud/overseas cloud APIs pose reachability risk from mainland China, with no mirror/offline guidance.
Clean frontmatter (name/version/argument-hint); repo has AGPL-3.0, changesets versioning, RELEASING.md, CONTRIBUTING.md, SECURITY.md, a clear vulnerability-reporting channel, and tag-driven signed releases with auto-update visible in CI. Deducted for: the skill file itself lacks FAQ and known-issues disclosure; CLI/MCP naming is not fully aligned (cal done vs complete_todo); the '-y' hybrid flag naming is confusable with 'confirm'.
The core workflow (load project memory → targeted search → read full threads → cited summary → write back) is well-structured with clear marginal value over manually re-tracing history, plus useful failure semantics. Deducted for: output usability not verified statically; ask_history RAG quality and distillation features depend on local Ollama or a cloud key, so several capabilities are unavailable to unconfigured users; cost/benefit evidence is thin.
Real CI exists (ci.yml runs fmt/clippy/cargo test; build.yml builds and signs on three platforms), README documents ignored real-data smoke tests, and the MCP registry entry allows cross-corroboration; claims are largely separated from fact. Deducted for: CI tests target the Rust core, not this skill file's key paths (MCP orchestration, CLI fallback flow); no third-party execution evidence, hence below the static cap of 5.
- The skill requires a pre-indexed Callimachus install; without indexing, most tools return empty results and the skill degrades to plain search — users must complete indexing and semantic-index build themselves.
- record_decision/record_gotcha/cal agents persistently write Callimachus memory or modify AGENTS.md/CLAUDE.md in repos; verify the managed block does not clobber hand-maintained content.
- First-run semantic-index build downloads an embedding model from Hugging Face; Ollama Cloud/cloud LLMs are overseas services that may be slow or unreachable from mainland China; keyword-only search and local Ollama are unaffected.
- Publisher is an unverified independent developer (BetaBots LLC / Ari Shaller) under AGPL-3.0 with a commercial dual license; enterprises should assess AGPL compliance obligations.
What does this skill do, and when should you use it?
Callimachus is a local-first desktop app (Tauri 2 + Rust) that indexes all your AI coding-agent threads into one local SQLite store, with hybrid keyword (FTS5/BM25) plus on-device semantic (sqlite-vec) search. It ships an MCP server, a cal CLI, and a VS Code/Cursor extension that all read the same index. The recall skill teaches agents to check that history before redoing work: load a project's decisions, gotchas and open TODOs, and avoid re-litigating settled choices. Everything stays on your machine; API keys live in the OS keychain.
Indexes conversations from 11 coding agents (Claude Code, Codex, Cursor, Gemini CLI, Qwen Code, Goose, OpenCode, Continue, Cline, Roo Code, Kilo Code) into local SQLite; exposes 21 MCP tools (17 read, 4 write) including search_threads, project_memory, recall_decisions, recall_gotchas, find_prior_work, ask_history (cited RAG), threads_for_file, and check_decision; write tools touch only Callimachus's own memory (record_decision, record_gotcha, complete_todo, snapshot_session); also provides the cal CLI (search, ask, memory, files, check, remember) and a VS Code/Cursor sidebar.
- A developer using several AI coding tools who wants to ask "how did we solve X?" without digging through raw logs
- Starting work in a repo with prior history and wanting to load its settled decisions, known gotchas, and open TODOs first
- About to make a technical choice (e.g. switching vector indexes) and wanting to check whether it was already decided and why
- Before editing a file, finding which past sessions touched it and which commits they produced
- Claude Code users who want resumable session snapshots across context compaction or tool handoff
What are this skill's strengths and limitations?
- Unified local index across 11 agent tools; data never leaves the machine and keys stay in the OS keychain
- Hybrid retrieval (BM25 + on-device sqlite-vec semantics) plus file-path search and thread→commit linking
- MCP server, CLI, and VS Code extension share one index.db, with two-way writes so agents record decisions/gotchas back
- Distillation can run on local Ollama or a logged-in Claude Code/Codex CLI subscription with no API key; open source under AGPL-3.0 with a commercial option
- High-value features (decisions, gotchas, RAG) depend on opt-in LLM distillation; without it these fields are empty, which can be misread as "no history"
- Cline/Roo/Kilo threads are index-only (no Resume); Cursor threads lack a project path; "Open in CLI" is unsupported on Windows/Linux
- Search requires the installed desktop app or a local build — it cannot stand alone in a bare API environment
- Background indexing lags; very recent sessions may not be embedded yet (keyword-only search still finds them)
How do you install this skill?
1) Install the desktop app from GitHub Releases (macOS .dmg / Windows .msi / Linux .AppImage/.deb); on first launch run Reindex to index your sources, then Build semantic index. 2) Wire the MCP server into any MCP client: claude mcp add callimachus -- npx -y callimachus-mcp. 3) The skill ships with the app at apps/desktop/src-tauri/resources/recall/ (bundled at .claude/skills/recall). Step-by-step source-build instructions are not fully documented in the source; you'd use pnpm install && pnpm desktop:dev.
How do you use this skill?
Say something like "didn't we already solve this?" to your agent or invoke /recall [what to recall]. Agents prefer MCP tools: call project_memory first, then search_current_project to scope to the repo, search_threads(hybrid: true) for fuzzy global recall, or ask_history for a cited [thread N] answer; use find_prior_work / check_decision before starting or re-proposing anything. Without MCP, use the cal CLI: cal search "query" -y, cal ask "question", cal memory. Write back with cal remember decision "..." --because "rationale".