loci-wiki 0.1.0

A personal memory server: aspect-tagged raw sources, concept-graph retrieval, and an MCP surface for Claude Code.

loci

A personal memory server. You save what you read; loci tags each source with aspects (methodology, knowledge-graph, …), wires up concept edges (citations, wikilinks, co-aspect), and serves the result to Claude Code over MCP so the model can recall the parts of your library that actually matter for what you're writing.

No interpretation graph, no draft pipeline, no PageRank. Just:

save → tag → recall

If you want a deeper read, start at docs/getting-started.md and then docs/architecture.md.

Status

Single-user, local-first. SQLite is the source of truth; raw blobs are content-addressed on disk. Python 3.12+.

Install

# with uv (recommended — isolated environment)
uv tool install loci-wiki

# with pipx
pipx install loci-wiki

# or just curl-pipe the installer
curl -fsSL https://raw.githubusercontent.com/loci-knowledge/loci/main/install.sh | sh

The first scan downloads the embedding model (BAAI/bge-small-en-v1.5, ~130 MB) into ~/.loci/models/.

Quick start

# 1. First-run setup: writes ~/.loci/.env (provider keys) and ~/.loci/config.toml
loci config init

# 2. Create a project
loci project create my-research

# 3. Register loci with Claude Code (one-time, user-scope)
claude mcp add loci --transport stdio --scope user -- loci mcp

# 4. Bind a directory to your project (so MCP knows which slug to use)
cd ~/Documents/my-research
loci project bind my-research

# 5. Save sources directly from Claude Code
#    (or via CLI: `loci save https://arxiv.org/abs/1612.03975`)

# 6. Recall in Claude Code
#    `@loci:source://...` for a single resource
#    `loci_recall("how does PPR work")` for ranked chunks with reasons

All user data lives under ~/.loci/. Run loci doctor to see resolved paths.

CLI commands

loci config init                              # write ~/.loci/.env + config.toml
loci doctor                                   # show storage paths + active project
loci server                                   # HTTP API + worker on 127.0.0.1:7077
loci mcp                                      # MCP stdio server (for Claude Code)
loci worker                                   # background worker only

loci project create <slug>                    # interactive wizard
loci project list / info / bind / manage
loci current set <slug>                       # pin project for MCP sessions

loci workspace create / list / add-source / scan / link / unlink
loci scan <project>                           # scan all linked workspaces
loci use [workspace_slugs...]                 # bind workspaces for this session

loci save <url_or_path> [--folder] [--aspects]
loci recall "query" [--aspects ...] [-n 10]
loci aspects [resource_id] [--add ... --remove ... --list-vocab]

loci status [project]
loci export [project]
loci reset                                    # wipe everything

MCP surface (Claude Code)

tool	what it does
loci_save	ingest URL/file/text, propose folder + aspects via elicitation, write to DB
loci_recall	concept-expand + BM25/ANN over chunks with concept-graph rerank
loci_aspects	list/edit aspects on a resource (elicitation form for editing)
loci_browse	list resources with folder + top aspects, filterable
loci_context	project profile + counts + top aspects for the current session
loci_research	paper-search sub-agent (deferred to v1.1)

@-mentionable resources:

@loci:source://{resource_id}
@loci:folder://{folder_path}
@loci:aspect://{label}

Source layout

src/loci/
  ui/cli.py              entry point: loci.ui.cli:main
  api/                   FastAPI REST + WebSocket
    routes/              projects, workspaces, sources, aspects, jobs
  mcp/                   MCP server + project resolution
  graph/                 sources, aspects, concept_edges, projects, workspaces
  retrieve/              lex + vec + hyde + concept_expand + pipeline
  capture/               ingest, folder_suggest, aspect_suggest, link_parser
  ingest/                walk → hash → extract → chunk → embed
  jobs/                  queue + worker + handlers (classify_aspects, parse_links, …)
  embed/                 sentence-transformers wrapper
  llm/                   pydantic-ai wrapper
  db/                    schema.sql + connection.py
  config.py              Settings + ~/.loci/ paths

License

MIT.