Lorekeeper

Self-improving memory for AI agents. One command, no cloud, no config.

pip install lorekeeper-mcp && lorekeeper setup && lorekeeper
Your agent remembers across sessions — and the memory gets better, not just bigger. Local. No API keys. No sign-up. Free to run forever.

Why Lorekeeper

Every AI agent session starts blank. You re-explain context, re-state preferences, re-teach patterns — every single time.

Files like CLAUDE.md and .cursorrules help, but they're hand-maintained, can't search themselves, and grow stale. Cloud services work, but your session data leaves your machine and you're paying per API call. Libraries are powerful, but you're writing the integration yourself.

Lorekeeper is a different shape: a local MCP server you pip install once. It connects to your existing agents, stores memories in SQLite on your own disk, and starts improving with every session:

Agent uses a memory → rates it useful or not →
scores adjust automatically → weak memories decay →
strong memories surface more often → search gets sharper

A fresh install and a six-month-old install are genuinely different products. The longer you use it, the less noise you get — and the more your agents feel like they actually know your codebase.

Quick Start

3 minutes, zero configuration:

# 1. Install
pip install lorekeeper-mcp

# 2. Configure your agents (auto-detects Hermes, Claude Code, Cursor)
lorekeeper setup

# 3. Start the MCP server
lorekeeper

lorekeeper setup scans for installed agents and injects the MCP entry, agent prompt, and bundled skills automatically. Use --check to preview without writing. Run lorekeeper --help or lorekeeper --version to verify the install.

Then ask your agent:

"Remember that I prefer curl -vX GET for debugging endpoints."

It calls lore_remember → memory stored. Next session:

"What's my preferred debug command?"

It calls lore_search → memory retrieved. ✅

Full walkthrough → docs/quickstart.md

Features

What	How
Hybrid search	Semantic vectors + BM25 keyword + time-decay + usage frequency + memory score — all ranked by a weighted formula
Self-improving	`lore_update` feedback adjusts scores. Bad memories fade (<2 confidence + not useful → soft-delete). Good ones rise.
Auto-linking	New memories are automatically linked to their closest semantic neighbor. A lightweight knowledge graph forms without effort.
Duplicate detection	New inserts are checked against existing memories. Near-identical content is blocked (override with `force=true`).
Dashboard	Full web UI — browse, search, edit, delete. Seven tabs including backup/restore with dedup preview.
Universal MCP	Works with Claude Code, Cursor, Hermes, Copilot, OpenCode — any MCP-compatible agent.
Local-first	Your data stays on your machine. SQLite + LanceDB (or ChromaDB). No cloud dependency, no API keys.
Namespaces	Multiple agents share one store with isolated namespaces. Writes go to your namespace; reads include the shared pool.
Reflection	Agents auto-extract learnings from sessions. Discoveries and lessons become searchable memories.

Use Cases

Staying in context across sessions

You set up your auth layer on Monday. Wednesday, a different agent starts fresh with no idea. With Lorekeeper, it already knows the middleware path, the token format, and which test covers the edge case — because you told it once.

"Remember that our JWT uses jose middleware in src/middleware/auth.ts
 and refresh tokens expire in 7 days."
→ lore_remember stores it
→ Every future session, on any agent: already in context

One memory pool, multiple agents

Claude Code for review, Cursor for implementation, Hermes for planning — they shouldn't each start from zero. Lorekeeper namespaces let them share one store. One agent's discovery becomes every agent's knowledge.

→ Engineer agent notes a performance quirk in the search service
→ lore_remember stores it under the shared namespace
→ PM agent surfaces it during sprint planning
→ No briefing, no copy-paste

Cross-session debugging

You fixed a subtle CORS bug three weeks ago and your agent helped. Neither of you remembers the details.

"What did we figure out about the CORS issue?"
→ lore_search returns the root cause, the fix, and the context around it

Project onboarding

New repo, new agent session. Instead of re-explaining the architecture, you run a few lore_remember calls after the first session. The next session — and every agent after it — starts with the right foundation.

"Remember: the payment service requires X-Idempotency-Key on all POST requests."
→ Claude Code, Cursor, and Codex all read from the same store

Who It's For

You, if you use:

Claude Code and want it to remember project context between sessions
Cursor and want persistent agent memory
Hermes, OpenCode, Codex CLI, Copilot CLI, or any MCP-compatible agent
Multiple agents and want them to share knowledge

Not for you yet, if:

You need team RBAC, audit logs, or SSO (coming post-beta)
You're building a consumer AI app (we're agent-first, not API-first)
You don't use AI coding agents (Lorekeeper is an MCP tool)

How It Compares

There are great tools in this space — each makes different trade-offs. Here's where Lorekeeper sits:

	File-based	Cloud services	Docker servers	Library (Mem0)	agentmemory (Node)	Lorekeeper
Setup	Built-in	API key + cloud config	`docker compose`	Write integration code	`npx agentmemory`	`pip install`
Data	Local	Cloud	Local	Your call	Local	Local
Search	grep	Vector	Vector	Vector	Hybrid	Hybrid + time-decay + usage + score
Self-improving	❌	❌	❌	❌	❌	✅ Quality loop
Knowledge graph	❌	Paid	❌	Paid	❌	✅ Free auto-link
Dashboard	❌	✅	❌	❌	Viewer	✅ Full web UI
Dependencies	None	~300MB	~2GB	~1.4GB	~200MB	~1.4GB (embeddings)
Built by agents	❌	❌	❌	❌	❌	✅ Dogfooded daily

Cloud services and Docker-based solutions are strong choices for teams or production apps. Lorekeeper is optimised for the other end: solo developers and agent workflows where zero ops, zero cloud, and a self-improving store matter most.

Dependency note: ~1.4GB is from the sentence-transformers embedding model (PyTorch). This is the same weight class as any local embedding solution. We're honest about it.

MCP Tools

Lorekeeper exposes 8 MCP tools covering the full memory lifecycle:

Tool	Purpose
`lore_search`	Hybrid semantic + keyword search with relevance scores
`lore_remember`	Fast one-shot memory save (auto-titles, auto-links)
`lore_insert`	Bulk structured insert with custom scores and links
`lore_update`	Feedback loop — rate memories, drive quality
`lore_forget`	Soft-delete wrong or outdated memories
`lore_reflect`	End-of-session: extract learnings, auto-save discoveries
`lore_processed_sessions`	Check which sessions are already processed
`lore_recommend_links`	Suggest candidate links between related memories

Full API reference → docs/api-reference.md

Performance

📊 Benchmark results from LKPR-70 coming soon. Lorekeeper is being benchmarked against agentmemory's published evaluation suite — retrieval precision, context token savings, and search latency at scale. Results will be published here once complete.

Early internal numbers on a 1,000-memory store:

Search latency: ~40–80ms (hybrid pipeline, LanceDB backend)
Duplicate detection: ~30ms per insert (0.6·semantic + 0.4·keyword threshold)
BM25 index rebuild: ~10ms at 5k memories

Full benchmark methodology and reproducible script → LKPR-70

Dashboard

A local web UI to browse, search, edit, and manage your memory store.

lorekeeper-dashboard
# → http://127.0.0.1:7777

Seven tabs:

Tab	What it does
Memories	Sortable table with live filter — title, score, confidence, usage, dates
Detail	Edit a memory's content, manage its links, soft-delete or hard-delete
Links	Browse the knowledge graph — source → relation → target
Query	Ad-hoc semantic + keyword searches with per-result score breakdown
Sessions	All processed agent sessions with extracted learnings
Config	Live tuning of search weights, quality thresholds, limits
Backup	Export/import memories as JSON with dedup preview

Lorekeeper Query tab — hybrid search with scores

Built by Agents, For Agents

Lorekeeper is developed using AI agents — Claude Code, Hermes, and our own agent team. The development cycle is itself a working demo of what it does:

agent builds a feature → uses Lorekeeper to capture what it learned →
searches those memories next session →
builds the next feature with the context already there

This isn't a marketing line. Every tool schema, return type, and workflow in Lorekeeper was shaped by agents using it daily — not by humans reading specs. When something was annoying to use, we changed it. When search returned noise, we tuned the weights. The product is what it is because the agents that build it depend on it.

The agentic development loop documented in this repo is how we actually work — and it's what Lorekeeper is designed to support for you.

Setup (Git clone)

If you prefer to clone the repo instead of pip install (e.g., for development):

git clone https://github.com/Jessinra/Lorekeeper.git
cd Lorekeeper
bash scripts/setup.sh

The setup script:

Installs dependencies with uv sync
Creates ~/.lorekeeper/ data directory
Installs git hooks (lint + tests enforced on commit)
Registers MCP config in detected agents (Claude Code, Cursor, Hermes)
Installs 5 agent skills for automatic memory workflows

Prerequisites: Python 3.11+, uv (or pip)

Development

# Unit tests (fast — excludes E2E)
uv run pytest

# E2E dashboard tests (requires Playwright — ~30s, runs a live browser)
uv run playwright install chromium  # one-time
uv run pytest tests/e2e/ -m e2e

# Lint
uv run ruff check src tests

# Dashboard dev
uv sync --extra dashboard
uv run lorekeeper-dashboard

# Type check (before push)
uv run mypy src

Full development guide → README.md Development section (below).

Project Layout

src/lorekeeper/
├── __main__.py          # Entrypoint — init_service() + mcp.run(stdio)
├── server.py            # FastMCP tool definitions (8 tools)
├── config.py            # Settings (pydantic-settings, LORE_ prefix)
├── models.py            # Pydantic models
├── dashboard/           # Web UI (FastAPI + uvicorn)
└── services/
    ├── orchestrator.py  # MemoryService — coordinates all sub-services
    ├── memory_engine.py # Vector store abstraction
    ├── lancedb_engine.py# LanceDB backend
    ├── chromadb_engine.py# ChromaDB backend
    ├── link_store.py    # SQLite — memories, links, reflections
    ├── keyword_index.py # BM25 index
    ├── search.py        # Hybrid ranking
    ├── dedup.py         # Duplicate detection
    ├── feedback.py      # Score delta, EMA, soft-delete
    ├── link_candidate.py# Link candidate pipeline
    └── relation_classifier.py # LLM-based relation classifier

Performance

All 500 LongMemEval-S questions, default hybrid weights (sem=0.45, kw=0.30):

Metric	Value	Latency
R@1	84.6%	32.9 ms/query
R@3	93.6%
R@5	96.6%
R@10	98.8%

Full per-category breakdown → docs/research/2026-06-11-retrieval-benchmark-results.md

License

Apache-2.0 — see LICENSE.

Built by agents, for agents. Manifesto · Strategy