lorekeep 0.1.0

Temporal knowledge graph for AI agents via MCP

Lorekeep

A temporal knowledge graph for AI agents, served read-only over MCP.

Lorekeep compiles a team's raw documentation into a versioned, time-aware knowledge graph (facts.jsonl) and exposes it to coding agents (Claude Code, Cursor, Codex) through the Model Context Protocol — with per-namespace permission and zero servers to run.

It applies Andrej Karpathy's "LLM Knowledge Base" idea: raw docs are the source code, the compiled graph is the executable. Knowledge is processed once at compile time, not re-RAG'd on every query.

---

Why

Existing tools each miss part of what a team needs:

	file-based	temporal KG	compile step	team permission	MCP
Obsidian + MCP	✅	❌	❌	❌	✅
mcp-knowledge-graph	✅	❌	❌	❌ (local)	✅
mem0 / cognee	❌ (DB)	partial	❌	partial (DB)	✅

Lorekeep targets the gap: strictly file-based + temporal graph + compile-once + namespace-scoped permission + MCP — for team-level (not just single-user) knowledge.

Features

• Compile-only — a curator (human + LLM) builds the graph; agents only read. No write path, no concurrency hell, deterministic output.
• File-sovereign — facts.jsonl (one fact per line, sorted) is the single source of truth and the sync unit (git or S3). No binary store committed.
• Temporal — every fact carries valid_from/valid_to (half-open [from, to)); query "what was true at T", history, diffs.
• Namespace permission — facts are tagged ns from the directory tree (raw/<ns>/); agents scoped to namespaces; cross-namespace edges hidden unless both endpoints are visible. Deny-by-default.
• MCP, stdio-first — lorekeep serve exposes 8 read-only tools; lorekeep mcp add wires Claude Code / Cursor / Codex. No server process to babysit.
• Lazy-reload — lorekeep compile updates the graph; the MCP server auto-refreshes on the next query. Connect once, use forever.
• Provider-pluggable extraction — litellm (OpenAI / Anthropic / DashScope/Qwen / Ollama). Strict-privacy → Ollama, fully local.
• Tier-1 eval — extraction P/R/F1 vs a gold corpus, entity-resolution F1, graph-structure metrics, determinism property tests.

Install

# from PyPI:
uvx lorekeep init                 # try it without installing

# or from a clone:
git clone https://github.com/manhhailua/lorekeep && cd lorekeep
uv tool install .                 # installs the `lorekeep` command

Quickstart

# 1. bootstrap a data home (~/.config/lorekeep + ~/.local/share/lorekeep)
uvx lorekeep init

# 2. add docs under the data home's raw/<namespace>/
mkdir -p ~/.local/share/lorekeep/raw/backend
cp your-docs.md ~/.local/share/lorekeep/raw/backend/

# 3. set a provider (edit ~/.config/lorekeep/config.yaml), then compile
uvx lorekeep compile                # raw/*.md -> graph/facts.jsonl

# 4. wire a coding agent (writes a portable .mcp.json)
uvx lorekeep mcp add --agent claude --ns backend

# 5. verify
uvx lorekeep doctor

Restart Claude Code → the 8 Lorekeep tools are available, scoped to your namespace.

How it works

                       COMPILE (offline, curator)                     SYNC
raw/<ns>/*.md ──► ingest ──► extract(LLM) ──► resolve ──► writer ──► facts.jsonl
                                                                            │
                                          ┌─────────────────────────────────┘
                                          ▼  (git pull / aws s3 sync)
                    SERVE + QUERY (runtime, per device)
facts.jsonl ──load──► GraphStore (networkx, temporal) ──► ScopedGraph (ns) ──► MCP ──► agent
                              ▲                                  │
                              └── lazy-reload on mtime change ◄──┘

Pipeline (ingest → extract → resolve → writer): markdown is chunked with provenance; an LLM extracts schema-constrained nodes/edges with temporal + namespace tags; aliases collapse to canonical entities; a deterministic writer emits sorted, byte-stable facts.jsonl + a manifest.json (provenance + errors + quarantine). Re-compiling unchanged input is byte-identical (per-chunk hash cache), so git diffs stay clean.

Serve: GraphStore loads facts.jsonl into a networkx graph with temporal queries. ScopedGraph is the single permission chokepoint — every query is filtered through strict visibility rules. The FastMCP server is a thin layer of read-only tools over ScopedGraph. It lazy-reloads when facts.jsonl changes, so compile is instantly visible without reconnecting.

Concepts

fact — one line of facts.jsonl, a node or edge:

{"kind":"node","id":"svc:payments","type":"service","ns":["backend"],"valid_from":"2024-01-15","valid_to":null,"props":{"lang":"go"},"src":["raw/backend/payments.md:12"]}
{"kind":"edge","id":"e_depends_on_0001","type":"depends_on","from":"svc:payments","to":"svc:auth","ns":["backend"],"valid_from":"2024-01-15","valid_to":"2025-03-01","props":{},"src":["...:20"]}

• ns — namespace set; ["public"] is globally visible.
• valid_to: null ⇒ current. History = multiple edges, same endpoints, different windows.
• src — provenance to raw doc line (audit, incremental re-compile, agent citations).

Permission — effective_ns = allowed ∪ {public}. Node visible iff ns ∩ effective_ns ≠ ∅. Edge visible iff both endpoints visible and edge.ns ∩ effective_ns ≠ ∅. Deny-by-default; an edge never reveals a neighbor the caller can't see.

Temporal queries — at_time(T) (snapshot of facts valid at T, half-open [from,to)), history(id) (versions of an entity), changes(t1,t2) (edges that began/ended in the window).

MCP tools (read-only, scoped)

search · get_node · neighbors · at_time · history · changes · list_namespaces · schema. Every result is filtered to the caller's namespace.

Configuration

config.yaml (resolved by precedence: explicit LOREKEEP_* env > LOREKEEP_HOME > dev marker > XDG):

provider:
  model: openai/qwen-plus                              # litellm model string
  api_base: https://dashscope-intl.aliyuncs.com/compatible-mode/v1
  api_key_env: DASHSCOPE_API_KEY                       # env var name (preferred)
  api_key: null                                        # or inline (gitignored config only)
ns:
  default: [public]
install_source: pypi                                   # pypi = portable .mcp.json

API keys never live in committed files — use api_key_env (env) or inline api_key in the gitignored config only. Examples (DashScope / OpenAI / Ollama) in .lorekeep/config.yaml.example.

Data home & dev mode

Path resolution (high → low): explicit LOREKEEP_* env → LOREKEEP_HOME → dev mode (.lorekeep/ or raw/ in CWD; auto-detected in a source checkout) → XDG (~/.config/lorekeep, ~/.local/share/lorekeep). So:

• Installed: uvx lorekeep init bootstraps the XDG home.
• Local dev: from the repo, uv run lorekeep compile uses the repo's raw/ + graph/ (zero migration).
• Custom KB: LOREKEEP_HOME=~/kb-work uvx lorekeep ….

See docs/compile.md and docs/serve.md.

Evaluation

Tier-1 (CI): extraction P/R/F1 vs a gold corpus, entity-resolution pairwise F1, graph-structure metrics, determinism. Run: uvx lorekeep eval. The north star is systematic thinking with complete information — memory-recall benchmarks (LoCoMo, LongMemEval) are parity checks, not the optimization target. See the design spec §16.

Project layout

src/lorekeep/
  models.py            shared contract (Node/Edge/Schema/Manifest)
  facts_io.py          facts.jsonl loader (store + eval)
  paths.py             4-tier path resolution (env/home/dev/XDG)
  defaults.py          default schema + config (for `init`)
  config.py, schema_io.py
  compile/{ingest,extract,resolve,writer}.py    the compile pipeline
  compile/providers.py                          LLMProvider (Fake/LiteLLM)
  store/{graph,fts}.py                          GraphStore + optional FTS cache
  perm/ns.py                                    ScopedGraph permission chokepoint
  mcp_server.py                                 FastMCP + 8 read tools (lazy-reload)
  integrations/{claude_code,cursor,codex,common}.py
  pipeline.py, cli.py
  eval/{gold,construction,retrieval}.py
tests/                 ~106 tests
docs/                  compile.md, serve.md, specs/, plans/

Status

v1 — compile pipeline + serve (store/permission/MCP/integrations) + data-home

• dev mode + lazy-reload, all merged to main, 114 tests green. Published to PyPI as lorekeep.

Roadmap (phase 2+): streamable-HTTP team server, OIDC/SSO, embeddings/hybrid search, wiki.md views, full Tier-2 benchmark datasets (HotpotQA/CronQuestions) and the bespoke Tier-3 Lorekeep-Reason eval.

Documentation

• Compile quickstart
• Serve to coding agents
• Design spec (architecture, permission, temporal, eval)
• Implementation plans: A compile, B serve, C data-home

License

Lorekeep is released under the MIT License — see LICENSE.

Copyright © 2026 Manh Pham. You're free to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the software, provided the copyright and permission notice are included in all copies. The software is provided "as is", without warranty of any kind.