Lorekeep

A file-sovereign, temporal knowledge graph shared by you and your coding agents over MCP.

Lorekeep compiles Markdown from multiple namespaces into a deterministic facts.jsonl graph, projects that graph into a human-readable Obsidian/Tolaria wiki, and exposes a compact namespace-scoped MCP surface to Claude Code, Cursor, Codex, and opencode. Agents can also propose facts during a session; proposals land in append-only journals and become visible after a confidence-gated resolve.

The LLM work happens during compile or an explicitly requested deep import. Ordinary graph queries, journal writes, resolve, wiki generation, lint, and status checks do not call another LLM.

What is available today

Lorekeep is suitable for one person using several coding agents and several devices, with an important current constraint: Git backup/sync is sequential. Simultaneous edits to the same raw document or journal can still require manual conflict resolution. A shared authenticated team server is roadmap work, not a shipped capability.

Get started

One-liner install (no uv needed — uses pipx or pip):

curl -fsSL https://raw.githubusercontent.com/manhhailua/lorekeep/main/scripts/install.sh | bash
lorekeep init

That's it. init sets up config, schema, provider, agent wiring, compiles any existing markdown, and starts a background daemon that auto-compiles future changes. Open the wiki in Obsidian/Tolaria and watch pages appear as the compile finishes.

Or with uv:

uv tool install lorekeep
lorekeep init

Non-interactive (CI, scripts):

lorekeep init --yes

What init does

Configures provider + namespace → writes about.md + profile.md → detects and wires coding agents (Claude Code, Cursor, Codex, opencode) → quick-imports available memory files → compiles if a provider key exists → starts the daemon. Idempotent — re-run anytime to pick up newly installed agents.

Add documents

# Drop Markdown under raw/<namespace>/
cp your-docs.md ~/.local/share/lorekeep/raw/backend/

# Compile runs in the background by default — wiki updates automatically
lorekeep compile

# Validate graph, schema, MCP, and provider
lorekeep doctor

The daemon watches raw/ and auto-compiles on file changes. No need to run compile manually unless you want immediate results.

Make the daemon persistent (survives reboot)

lorekeep agent service install   # systemd (Linux) / launchd (macOS)

Upgrade

lorekeep update          # upgrade to latest from PyPI
lorekeep update --check  # preview without upgrading

Wire an agent that wasn't auto-detected

lorekeep agent wire --agent codex --scope user --ns backend

Restart the agent after its MCP config changes. Open the wiki with lorekeep wiki --open.

Runtime model

COMPILE / CURATE

raw/<ns>/*.md ──> chunk ──> extract(LLM, cached) ──> resolve ──┐
                                                               │
pending/<ns>/journal.jsonl ──> confidence gate + replay ────────┤
                                                               v
                                         facts.jsonl + manifest.json
                                                   │
                                                   ├──> wiki/*.md
                                                   └──> local FTS cache

SERVE / USE

facts.jsonl ──> GraphStore ──> ScopedGraph(allowed namespaces) ──> MCP
     ^                    lazy reload on facts.jsonl mtime             │
     └──────────── resolve <──── namespace journal <──── agent write ─┘

Raw Markdown, schema.json, and accepted/pending journals are the durable knowledge inputs. The graph, manifest, wiki, cache, and FTS index are derived and can be rebuilt on each device.

Compile, resolve, and wiki

lorekeep compile is the normal all-in-one operation:

1. chunk raw/ with path:line provenance;
2. extract typed nodes, edges, aliases, summaries, and relation descriptions;
3. resolve aliases, validate facts, and quarantine invalid candidates;
4. write sorted facts.jsonl and manifest.json atomically;
5. replay/merge journals when present; and
6. generate the wiki once from the final graph.

Unchanged chunks use a hash cache, so they do not repeat extraction calls; sorted publication keeps the resulting graph byte-stable for unchanged inputs.

Use the narrower commands when only the derived view or journals changed:

lorekeep resolve       # merge pending journal entries; zero LLM calls
lorekeep wiki --open   # re-project the existing graph; zero LLM calls

Daemon and service

lorekeep agent watch polls at a configurable interval (60 seconds by default) and currently performs event-driven maintenance:

• raw file count/mtime or schema change → compile;
• journal mtime change → resolve;
• Claude/Codex memory change → quick import;
• supported live transcripts → bounded Markdown dumps under raw/;
• detected agent change → idempotent MCP/hook wiring;
• successful compile → self-heal, wiki refresh, and backup sync when configured;
• installed Lorekeep version change → restart the running watcher.

It does not currently run nightly lint, weekly suggestions, or an autonomous schema-evolution scheduler. Run those one-shot operations explicitly:

lorekeep agent lint
lorekeep agent lint --auto-fix
lorekeep agent suggest
lorekeep agent status

For login/restart persistence:

lorekeep agent service install
lorekeep agent service status

MCP contract

The runtime exposes exactly seven composable tools:

Clients that support MCP resources can also read:

• lorekeep://schema
• lorekeep://namespaces
• lorekeep://status

Every graph-fact query and graph statistic goes through ScopedGraph. Effective visibility is the configured scope plus public; an edge is returned only when its own namespace and both endpoints are visible. Static schema and aggregate compile/pending operational metadata are process-wide. Write namespaces come from the verified server scope, not from caller payloads.

Making the MCP server available does not guarantee that every coding agent will choose to call it. init and mcp add print an instruction snippet that tells the agent to use this retrieval sequence:

context(section="status") → search(query) → get_node(id) → neighbors/temporal_query

Keep that snippet in the agent's project/user instructions when the client does not persist it automatically. Agents should cite src, check graph freshness, and treat “not found” as “absent or outside this namespace,” not proof that a fact does not exist globally.

Configuration

All model names must use LiteLLM's {provider}/{model} form. Native providers include OpenAI, Anthropic, DeepSeek, DashScope/Qwen, Gemini, OpenRouter, Mistral, Groq, Together AI, and others exposed by LiteLLM. Ollama, vLLM, and LM Studio are available for local/custom endpoints.

provider:
  model: deepseek/deepseek-chat
  api_key_env: DEEPSEEK_API_KEY
  timeout_seconds: 120
  max_retries: 2
compile:
  language: en
ns:
  default: [me]
  personal: me
agents:
  enabled: [claude, codex, cursor, opencode]
  auto_wire: true
  wire_scope: user
  watch_transcripts: true
  self_heal: true

Prefer provider.api_key_env. An inline provider.api_key is accepted only in the local gitignored config.yaml. Native providers normally need no api_base; set it for Ollama on a non-default host or another custom OpenAI-compatible endpoint. See the validated configuration example.

Change settings without editing YAML:

lorekeep config show
lorekeep config set provider.model openrouter/deepseek/deepseek-chat
lorekeep config set provider.api_key_env OPENROUTER_API_KEY
lorekeep config set compile.language vi
lorekeep config set ns.default me,backend
lorekeep config set agents.wire_scope user

Optional LiteLLM tracing is available through Langfuse or LangSmith by setting observability.provider and the corresponding environment credentials.

compile.language is a lowercase ISO 639-1 code, defaults to en, and keeps LLM-extracted names, summaries, and descriptions consistent even when source files mix languages. Changing it invalidates the relevant extraction cache entries on the next compile. Raw Markdown, proper nouns, stable IDs, and technical identifiers are preserved.

Data home and paths

Path precedence, high to low:

1. per-path LOREKEEP_RAW, LOREKEEP_OUT, LOREKEEP_CACHE, LOREKEEP_SCHEMA, LOREKEEP_CONFIG, LOREKEEP_PENDING, LOREKEEP_WIKI, or LOREKEEP_LOGS;
2. LOREKEEP_HOME;
3. development mode (.lorekeep/ in the current checkout or LOREKEEP_DEV=1); and
4. default dotdir ~/.lorekeep/.

All platforms (Linux, macOS, Windows) default to ~/.lorekeep/ for both config and data. See the data-home guide for other platforms and overrides.

Multi-device backup

Lorekeep backs up to a private Git remote — durable inputs (raw markdown, schema, journals) plus the latest graph/wiki snapshot:

# One-time setup (create a private repo first on GitHub)
lorekeep backup --init https://github.com/<you>/lorekeep-data.git

After that, the daemon auto-syncs after every compile, resolve, and self-heal. Manual sync:

lorekeep backup

Restore on a new device:

# Install lorekeep, then clone the backup into the data home:
git clone https://github.com/<you>/lorekeep-data.git ~/.lorekeep
lorekeep init --yes    # creates local config, rewires agents, preserves data

The restored graph and wiki are immediately usable — no recompile needed. Config/secrets, cache, FTS, logs, and Obsidian settings stay local and are never backed up.

Generated graph/wiki files are marked non-mergeable: Git may merge durable inputs but never silently combines two compiled snapshots. If both devices changed durable sources, reconcile and compile once.

Diagnostics and support

lorekeep doctor is the pass/fail installation check. Runtime logs live under the resolved logs/ directory and avoid prompts, raw docs, fact properties, journal payloads, and credentials.

lorekeep doctor
lorekeep agent service status
lorekeep support                 # print report + create redacted ZIP
lorekeep support --report-only
lorekeep support status          # automatic issue-reporting state
lorekeep support off             # disable automatic issue creation

The support bundle contains an allowlisted report, redacted log tail, and manifest counters—not raw knowledge or configuration. See runtime logging and bug reports.

Current limits

• Search is keyword/FTS plus graph traversal; hybrid/vector retrieval is planned.
• Git sync is sequential; conflicting simultaneous edits need manual resolution.
• Local stdio is the supported transport; authenticated shared-team HTTP hosting and OIDC/SSO are not shipped.
• Lint, suggest, and contribution analysis are one-shot commands, not scheduled jobs.
• There are no built-in software-source connectors for repositories, observability systems, CI, Confluence, PDFs, or URLs yet.
• Coding-agent tool use depends on the client's MCP support and instructions.

See the roadmap for unshipped directions. Architecture docs describe only current behavior unless a section is explicitly marked planned.

Development

git clone https://github.com/manhhailua/lorekeep.git
cd lorekeep
uv sync
uv run lorekeep init          # uses .lorekeep/ in the repo (dev mode)
uv run pytest
uv run pytest tests/test_core_regression.py -q
uv run python scripts/generate_cli_reference.py --check
uv build

Tests use FakeProvider; no API key or network call is required. Determinism and the compact seven-tool MCP surface are regression contracts.

Documentation

Start at the documentation index.

• Getting started
• Compiling and resolving
• Importing agent sessions
• Serving the graph over MCP
• Browsing the wiki
• Backing up and syncing
• Runtime logging and bug reports
• Architecture overview
• Generated CLI reference

License

Lorekeep is released under the MIT License.