shelfmark

Local hybrid retrieval for agent memory and codebases. One SQLite file, no services, no cloud. BM25 + dense embeddings fused with Reciprocal Rank Fusion, an optional cross-encoder reranker, and an MCP server so coding agents (Claude Code, Codex, anything MCP) can recall your notes, decisions, docs, and code instead of guessing.

A shelfmark is the code a librarian writes on a book so it can be found again. That is the whole product: before your agent answers, the librarian fetches the handful of notes and code chunks most likely to matter and puts them on the desk.

Why this exists · Quickstart · MCP server · What gets indexed · Evaluation · How it compares · Configuration

Why this exists

Agent harnesses accumulate knowledge: memory notes, ADRs, plans, standards, handoffs, and the code itself. The useful answer to most questions is already written down somewhere. shelfmark indexes all of it into one local SQLite file and answers "what did we decide about X / where did we handle Y" in a single query, with path:line citations.

shelfmark is the retrieval engine that came out of a year of running AI coding agents daily — most recently distilled from forgekit, an AI dev toolkit for coding agents. Every design choice below was forced by a real failure mode in that daily use, not picked off a paper.

Design choices that fell out of a year of measured iteration (see eval/holdout-policy.md and inline rationale comments):

• Hybrid by default. BM25 catches identifiers and exact terms; embeddings catch paraphrase. RRF fusion beats either alone on mixed corpora.
• Code-aware tokenization. camelCase/snake_case sub-tokens on both corpus and query sides ("create player" matches createPlayer). Measured +2.8pp code / +3.2pp overall.
• Symbol-definition boost. Chunks whose defined symbol matches a query identifier get a rank-0 signal (+2.8pp code, zero regressions).
• Selective reranking. Cross-encoder rerank helps code and standards, hurts memory recall (measured -10.5pp). The rerank policy is scope-aware; memory is never reranked.
• Contextual chunk prefixes. Every chunk embeds with a type | repo | file | symbol header, not raw text.
• Freshness without wall-clock. Optional recency prior for memory scope, deterministic (reference = max mtime among candidates), so evals stay reproducible.

Quickstart

pip install shelfmark-rag   # or: pipx install shelfmark-rag

shelfmark-build      # first run writes a starter ~/.shelfmark/sources.yaml — edit it, then rerun
shelfmark-query "how do we handle retry timeouts"

First build downloads intfloat/multilingual-e5-small (~120MB). Everything after that runs offline.

git clone https://github.com/LucasSantana-Dev/shelfmark && cd shelfmark
python3 -m venv venv && venv/bin/pip install -e .

venv/bin/shelfmark-build
venv/bin/shelfmark-query "how do we handle retry timeouts"

MCP server (agent integration)

// e.g. Claude Code: .mcp.json or ~/.claude.json
{
  "mcpServers": {
    "shelfmark": {
      "command": "shelfmark-mcp"
    }
  }
}

(Local checkout instead of pipx install? Use "command": "/path/to/shelfmark/venv/bin/shelfmark-mcp".)

Two tools:

• rag_query — full-corpus hybrid search (code, docs, commits, notes). Auto-scopes to the repo your agent is working in.
• search_knowledge — cross-project search over durable knowledge only (memory/standards/plans/handoffs/adrs; configurable via RAG_KNOWLEDGE_SCOPE). Never reranked, by measurement.

examples/claude-code/ has the full loop: auto-recall on every prompt (UserPromptSubmit), incremental reindex on file writes (PostToolUse), drift reindex + weekly report at session start, and a nightly rebuild with an eval regression gate.

What gets indexed

sources.yaml declares everything (see sources.yaml.example):

Incremental reindex (indexer.py --incremental <files>) keeps writes cheap; session_chunker.py can additionally index agent session transcripts.

Evaluation

The eval harness is the part most RAG setups skip. eval/run.py scores Hit@1/3/5 + MRR per scope against a JSONL dataset; eval/check.sh gates any change at >5pp regression vs a frozen baseline; eval/holdout-policy.md documents the train/holdout discipline (the holdout set is never used for tuning — numbers quoted from it are honest).

This repo ships a public, reproducible dataset (eval/dataset-public.jsonl) whose queries target this repository's own code and docs:

venv/bin/python indexer.py                 # index this repo (sources.yaml.example works as-is)
venv/bin/python eval/run.py --dataset eval/dataset-public.jsonl --label mine

Benchmark results and methodology: BENCHMARK.md.

To evaluate on YOUR corpus, write ~50 {"query", "expect_path_contains", "expect_scope"} lines, freeze a fifth of them as holdout, and wire eval/check.sh into your nightly rebuild. That regression gate is what keeps retrieval quality from silently rotting as the corpus grows.

How it compares

No cross-tool benchmark exists yet (see Methodology & honest limitations — we'd genuinely like to see one run). Qualitative trade-offs, no fabricated numbers:

Pick a hosted vector DB when you need multi-tenant scale across millions of documents — that's a different problem than agent recall over your own repos.

Want to measure and improve retrieval ranking quality on your own pipeline, independent of any specific agent? hitgate provides label-free regression testing for hybrid retrievers. shelfmark and hitgate share a common hybrid-retrieval foundation (BM25 + embeddings + RRF) but serve complementary use cases: shelfmark for zero-setup agent memory with MCP integration, hitgate for ranking evaluation and quality gates on any retriever you already have.

Configuration

All optional — see .env.example for the full list. Highlights:

Client layers

If you work for more than one client, keep each client's business knowledge out of the index every session reads. Declare clients in sources.yaml:

clients:
  acme:
    roots: [~/dev/acme-app, ~/notes/acme]

• Files under a client's roots, or with client: acme in their frontmatter (top level or under metadata:), are indexed into that client's own file (index.client-acme.sqlite), never the general index. client: none keeps a note general even inside a root (logged at index time). Once a client key is present, anything else (unknown slug, list, empty, broken YAML) skips the file instead of sending it to general.
• A query reads the general index plus the active client's file only. The active client comes from the process (RAG_CLIENT, or the process cwd under a client root), never from a tool-call argument, so an agent cannot switch itself into another client's layer.
• Session transcripts are routed by the cwd they recorded. Known limit: a session started outside every client root that then reads a client's files is indexed as general. Start client sessions inside the client's root, or leave session_chunker.py off.
• BM25 is scored per file, so one client's vocabulary never shapes another's term statistics; results are then ranked together. With no clients declared, ranking is unchanged.
• A client's data lives only in its file and its index.client-<slug>.backup-* snapshots. Removing a client from sources.yaml while its file still exists stops the next build (its untagged files would otherwise land in general): archive or remove the file and its sources globs first. The last build's clients are recorded in $RAG_HOME/clients.json, so this also holds for a client with a custom db: path.
• Upgrading from a version without client layers: run one full rebuild. Chunks now store the resolved path, and rows written before keep the old spelling until rebuilt.
• Root matching ignores case on macOS/Windows (str.casefold, close to but not exactly the file system's own folding).

To offboard a client, purge before removing it from sources.yaml:

shelfmark-purge acme --lexicon acme-terms.txt                       # dry run: what would go
shelfmark-purge acme --apply --archive age1... --lexicon acme-terms.txt

The purge streams the client's index, its backups and its query-log rows into one age-encrypted archive (never plaintext on disk). Then it deletes them and scrubs any residue of the client from the general index and its backups with secure_delete + VACUUM, so nothing survives in free pages or the WAL. It then checks that the lexicon's canary terms appear nowhere in the raw bytes of what remains. Notes under the client's roots tagged client: none (lessons you decided are general) are kept. --no-archive deletes without an archive. A tombstone keeps later builds from routing the client's files into general if its globs stay. Source files are never touched.

Contributing

Issues and PRs welcome — especially a real cross-tool benchmark (see How it compares), support for more embedding models, or non-Claude MCP client examples. If shelfmark saves your agent a guess, a star helps others find it.

License

MIT