scry

Self-hosted semantic code search and code-anchored memory, shared across your devices. One scry serve instance holds a hybrid dense + BM25 index in a single SQLite file; every checkout of the same repo, on any machine, syncs and searches the same index. No third-party service ever sees your code.

scry "how are chunks defined"           # semantic search in the current repo
scry "auth middleware" src/server       # scoped to a subdirectory
scry --repo github.com/you/proj "query" # any indexed repo, from anywhere
scry -a "how does sync decide what to upload"   # cited answer from a local LLM
scry --web --answer "sqlite-vec quantization"   # web-grounded answer (Tavily)
scry watch                              # keep the index in sync while you work
scry remember "lesson learned" --kind lesson --pain 7 --anchor src/store.rs
scry recall "why did we pick sqlite"    # memories about this codebase

Outside a project (your home directory, say), scry "query" searches every indexed repo and prefixes each hit with its repo key.

Docs: https://verdenroz.github.io/scry/

How it works

• Repo-keyed index. Repos are identified by their normalized git remote (fallback: .scry.toml name, then directory name) and paths are stored repo-relative, so a second checkout of an indexed repo syncs with zero uploads. Sync is a manifest diff by xxh64 hash; only changed files travel, and chunks whose content is unchanged reuse their vectors.
• Structure-aware chunks. Tree-sitter definition-level chunks with symbol paths for 16 languages (Rust, Python, TS/JS, Go, Java, C, C++, C#, Kotlin, Ruby, PHP, Scala, Lua, Bash, ...), blank-line-snapped windows for everything else. Each chunk embeds with a repo > path > symbol context header.
• Routed hybrid retrieval. Dense KNN (sqlite-vec, cosine; binary coarse pass + float rescore past ~8k chunks) fused with FTS5 BM25 via weighted RRF. Keyword-shaped queries lean lexical, natural-language questions lean dense; the BM25 leg is expanded with fuzzy matches from the repo's symbol table. Recency boost and near-duplicate filtering on top; optional HyDE expansion through the chat model.
• Memory. Typed notes (lesson, decision, convention, skill, fact, episode) with write-time salience (surprise + cost + explicit weight), recency decay with reinforcement on helpful recalls, and anchors: a memory tied to a file goes stale the moment a sync commits a different content hash, so it can never stay confidently wrong.
• Local models. Embeddings come from any OpenAI-compatible /v1/embeddings endpoint (llama.cpp, llama-swap, Ollama); --answer uses any OpenAI-compatible chat endpoint; --web uses Tavily. Each is optional and reports cleanly when unconfigured.

Quick start

cargo install --path crates/scry
mkdir -p ~/.config/scry && cp deploy/config.example.toml ~/.config/scry/config.toml
# edit: point [embedding] at your endpoint; add [chat]/[tavily] if wanted
printf 'TAVILY_API_KEY=...\n' > ~/.config/scry/env && chmod 600 ~/.config/scry/env

cp deploy/scry.service ~/.config/systemd/user/
systemctl --user enable --now scry
loginctl enable-linger $USER      # start at boot without a login

cd ~/projects/yourrepo && scry index
scry "where is the retry logic"

Secrets stay in ~/.config/scry/env (loaded by the unit, referenced from the config as "env:VAR"), so config.toml is safe to keep in dotfiles.

Other devices point [client] server_url (or SCRY_SERVER_URL) at the machine running scry serve, set the shared bearer token, and run scry index in their checkout; unchanged files upload nothing. See docs/hosting.md for LAN, WireGuard/Tailscale, docker compose (bundled llama.cpp embedder), and public-HTTPS setups.

Claude Code plugin

claude plugin marketplace add /path/to/scry/plugin
claude plugin install scry@scry

Per session it spawns scry watch in the project (SessionStart), injects relevant memories into each prompt (UserPromptSubmit), and cleans up on exit. In a directory that is not a project (your home directory, for example) it stands down entirely and Claude keeps its builtin tools.

Commands

Command	Purpose
scry "query" [path]	semantic search; -m N, -c content, -w web, -a answer
scry --repo <key> "query"	search one indexed repo from anywhere; no repo + no flag = all repos
scry serve	run the index server
scry index / scry watch	one-shot / continuous sync of the current repo
scry status	repos, files, chunks, memories on the server
scry remember / recall	write / retrieve memories (--kind, --pain, --anchor)
scry memory helpful|noise <id>	reinforce or demote a memory
scry repo prune <key> [--into <key>]	drop an index; migrate or detach its memories
scry eval cases.toml	Recall@10 / MRR against a golden query set

Development

Hot paths carry soothfast benchmarks; CI gates performance against the merge-base and checks doc claims against fresh measurements. cargo test --workspace runs the suite, cargo soothfast measure -p scry-core --save-baseline dev records a local baseline, scry eval guards retrieval quality when touching the ranking pipeline.