loci 🧠

Two thousand years ago, orators stored their speeches in the rooms of a
palace and walked through them to remember. loci does the same for your
files.

Loci is the method behind every memory palace: place knowledge in
locations, recall it by walking the path.

A queryable "second brain" for the project docs, notes, and chat logs scattered
across a dozen directories — and an MCP server so your AI agents can use it too.

Local files → heading-aware chunking → embeddings → hybrid retrieval (vector +
BM25) → LLM answer with section-level citations. The index lives entirely on
your machine; only embedding/chat calls go out, to any OpenAI-compatible API
(Zhipu / DeepSeek / Kimi / OpenAI / …).

The thesis (from studying the 90k-star platforms and the graveyard of
dead lightweight tools — see
our competitive landscape study):
don't build another chat app. Build the memory layer that every chat app
can mount. Claude Desktop, Cursor, Cline, or any MCP host becomes this
project's UI, for free.

Demo

Real session, indexed against the docs of
minimax-h3-turing
(paths shortened for display):

$ python main.py search "what the 22G card can and cannot do" -k 3

[1] minimax-h3-turing/docs/en/01-hardware-limits.md > 01 · What a 2080Ti 22G Can and Cannot Do    (similarity 0.562)
[2] minimax-h3-turing/docs/en/02-w4a8-vs-w4a4.md > 02 · Quantization Measured > You Can Try Without 22G  (similarity 0.446)
[3] minimax-h3-turing/docs/en/01-hardware-limits.md > ... > 3. VRAM is just barely enough — manage it  (similarity 0.504)

$ python main.py ask "How should I choose between T8 aggressive mode and the final-render mode, and why?"

Answer:
* Drafts / preview / shot selection: use T8 aggressive mode — a 43% speedup
  (2.7 min/clip), and "a different picture of equal quality" is fine for picking shots.
* Final shots: use final-render mode (no T8). T8 makes the numerical trajectory
  fork, so re-running with the same seed produces a different clip — which breaks
  the reproducibility final outputs need.

[source: docs/en/08-t8-blockcache-4step.md > Practical Advice (4-step Turbo route)]
[source: docs/en/06-faq.md > 12. Cache-style accelerators break "same-seed re-runs"]

Hybrid retrieval means a Chinese query still finds the English doc (and vice
versa) — keyword evidence (BM25) catches what embeddings miss, and every
citation points at a section, not just a file.

Does hybrid actually help? (mini-eval, 10 bilingual queries)

$ python scripts/eval_retrieval.py scripts/eval_cases.example.jsonl
vector-only: 9/10  →  hybrid: 10/10

Hybrid also fixed the #1 ranking on keyword-ish queries (e.g. "T8 block cache
threshold speedup": vector put an FAQ first, hybrid puts the actual T8
writeup first). Run it against your own corpus with your own cases file.

Reranking: two providers

--rerank reorders the fused candidates for precision:

python main.py search "T8 speedup" --rerank          # provider from config
python main.py search "T8 speedup" --rerank local    # cross-encoder (BAAI/bge-reranker-base)

The local model downloads on first use (~1.1 GB; set HF_ENDPOINT=https://hf-mirror.com
if HuggingFace is slow in your region). Measured on a 2080 Ti, bilingual query.

Office documents, PDF tables, chat logs

• PDFs: with the [pdf] extra, PyMuPDF4LLM extracts pages as markdown —
  tables come through as pipe rows (plain pypdf text is the fallback)
• Word: with the [docx] extra, .docx paragraphs and table rows are indexed
• Chat exports: drop a ChatGPT or Claude conversations.json into any
  source directory — it becomes one searchable document per conversation,
  tagged chatlog (search --tag chatlog scopes to chat history)

How it relates to Obsidian / your note app

It doesn't compete — the two layer up. Obsidian (or any editor) is the
note-taking frontend; this is the cross-vault search engine: point
sources at any directories (Obsidian vaults, project docs, chat exports)
and query all of them at once — from your terminal, your scripts, or your AI
agent via MCP. Obsidian-native details are understood: frontmatter tags:
(filter with search --tag), [[wikilinks]] (walk the graph with links),
code blocks are never cut mid-block, and one-line notes stay searchable.

How it works

flowchart LR
    subgraph sources["📥 Your machine"]
        notes["Obsidian / markdown notes"]
        docs["PDF tables · docx · project docs"]
        chats["ChatGPT / Claude exports"]
        mem["memories/ — agent-written notes"]
        wikidir["wiki/ — consolidated pages"]
    end

    subgraph loci["🧠 loci — local index, nothing leaves the machine"]
        ingest["ingest / watch<br>loaders → chunker → embedder"]
        store[("ChromaDB<br>hybrid index")]
        retrieve["hybrid retrieval<br>vector + BM25 → RRF"]
        mcp["loci-mcp<br>8 tools · resources · prompts"]
    end

    subgraph hosts["🖥️ Your AI hosts"]
        ide["Claude Code · Qoder · Trae<br>Cursor · Cline"]
        desktop["Claude Desktop"]
        term["Terminal<br>search / ask / chat / wiki"]
    end

    api["☁️ OpenAI-compatible API<br>Zhipu / DeepSeek / Kimi / OpenAI<br>or 100% offline via Ollama"]

    sources --> ingest --> store
    mem -. auto-indexed .-> store
    wikidir -. auto-indexed .-> store
    store --> retrieve
    retrieve --> term
    retrieve --> mcp
    mcp <--> ide
    mcp <-.-> desktop
    retrieve -. "embedding + chat calls only" .-> api

The write path in one line: loaders → chunker (heading-aware split) → embedder → store (ChromaDB, persistent) — incremental, deduplicated by content hash.

Install & quick start

Requires Python 3.11+ (uses the stdlib tomllib).

# option A: install as a package (adds `loci` and `loci-mcp` commands)
pip install -e ".[pdf,docx]"   # optional extras: PDF w/ tables, Word documents

# option B: zero-install quickstart
pip install -r requirements.txt

# 1. Configure: copy the example and fill in your values
cp config.example.toml config.toml

# 2. Ingest (incremental — deduplicated by content hash, safe to re-run)
loci ingest            # or: python main.py ingest

# 3. Ask
loci ask "what did I write about X?"

The workflow

flowchart TD
    A["pip install loci-rag"] --> B["cp config.example.toml config.toml<br>fill API keys + source dirs"]
    B --> C["loci ingest — hybrid index built"]
    C --> D["loci watch — index stays fresh (optional)"]
    C --> E{"What do you need?"}
    E -->|"a synthesized answer"| F["loci ask --verify<br>claim-by-claim audit"]
    E -->|"raw excerpts to quote"| G["loci search --tag memory"]
    E -->|"back-and-forth"| H["loci chat"]
    E -->|"scattered notes on a topic"| I["loci wiki topic<br>consolidate into a wiki page"]
    F --> J["loci remember —<br>keep what you learned"]
    I --> J

Commands

Filter operators (combine freely, on search and ask):

One memory, every IDE

Because every MCP host mounts the same loci server (same config.toml, same
index), memory written from one tool is recalled from every other:

# Claude Code
claude mcp add loci -- loci-mcp

// Cursor / Cline / Qoder / Trae (mcpServers JSON — same shape everywhere)
{ "mcpServers": { "loci": { "command": "loci-mcp" } } }

Then, from any of them: "remember that the staging password rotates on
Mondays" → brain_remember → later, from a different IDE:
"when does the staging password rotate?" → answered, with the memory cited.
Memories live as plain markdown in the memories directory (git-friendly, no
lock-in) and are tagged memory, so loci search --tag memory scopes to them.

Cross-IDE tip: the default store / memories paths are relative to the
directory loci is launched from. If your IDEs start in different project
folders, point both at one absolute location in config.toml — e.g.
store.path = "~/.loci/store" and memories.path = "~/.loci/memories" —
and every IDE shares the exact same memory store.

sequenceDiagram
    participant CC as Claude Code
    participant L as loci-mcp
    participant S as ChromaDB (local)
    participant T as Trae / Qoder / any IDE
    CC->>L: brain_remember("deploy rotates Mondays")
    L->>S: write memory.md + embed + index
    Note over S: persists across sessions and IDEs
    T->>L: brain_search("password rotation")
    L->>S: hybrid retrieval
    L-->>T: cited answer — the memory is recalled

Mount it in any MCP host

Add to claude_desktop_config.json (Claude Desktop) or your MCP client's
config:

{
  "mcpServers": {
    "loci": {
      "command": "python",
      "args": ["/path/to/loci/mcp_server.py"]
    }
  }
}

The server exposes three tools (zero dependencies beyond the core):

Beyond tools, the server speaks the full protocol:

• Resources — resources/list exposes brain://stats plus one
  brain://note/… resource per indexed file (raw markdown via resources/read)
• Prompts — three ready-made templates: brain-briefing, study-plan,
  contradiction-check; hosts render them with your topic pre-filled

Fully offline with Ollama

The index is local by design — and the embedding/chat calls can be too. Any
OpenAI-compatible server works; Ollama is verified
end-to-end:

[llm]
base_url = "http://localhost:11434/v1"
api_key = "ollama"          # any non-empty placeholder
model = "qwen2.5:0.5b"

[embed]
base_url = "http://localhost:11434/v1"
api_key = "ollama"
model = "all-minilm"

With this config, ingest / search / ask make zero cloud calls.
Swap in a bigger local chat model for better answers — the pipeline is
model-agnostic.

Configuration

API keys can also come from the environment variables BRAIN_LLM_API_KEY /
BRAIN_EMBED_API_KEY (these override the config file).

Design decisions

• ~300 lines of core, no LangChain — every stage is readable, hackable,
  and learnable. The whole engine fits in one sitting.
• MCP-first — the agent ecosystem is the UI layer. No web app to maintain.
• Hybrid retrieval on by default — vector search fused with a native
  ~60-line BM25 (CJK-aware tokenizer) via Reciprocal Rank Fusion.
• Citations always, with breadcrumbs — path > section, so claims are
  verifiable at a glance.
• Robust, inspectable indexing — defensive loaders (skip what can't be
  parsed, never hang), content-hash incrementality, real pruning, stats and
  doctor so the index is never a black box.
• Tiny notes stay searchable — no minimum-chunk filter; a one-line note is
  still indexed (a lesson from watching other tools drop or choke on them).
• Keys never in code — config.toml (gitignored) or env vars.

Where it sits

(Full data and reasoning: competitive landscape study.)

Roadmap

See docs/roadmap.md — reranking, GraphRAG experiments, more loaders.

License

MIT