RDF knowledge graph for LLMs — an MCP server that gives any AI agent a structured, queryable memory via six tools: triple, sparql, graph, search, validate, analyze.
Plug it into any project as a persistent brain. Use the built-in ontology for research notes, construction planning, sales pipelines, recipe collections, or any structured knowledge. Auto-installs Claude Code skills and hooks.
RDF knowledge graph for LLMs — an MCP server with 5 tools: triple, sparql, graph, search, validate
On first launch, Hippocamp automatically sets up Claude Code hooks and skills in your project — no manual configuration needed.
Why a graph?
Real test: the same question — "what's needed for electricity?" — asked two ways in a home construction project.
File scanning
Hippocamp
Time
~3 min (38 tool calls)
~2 sec (2 queries)
Tokens
60,500
~2,000
Result
Complete but verbose
Structured, concise
Context
Large text blocks
Compact JSON
15× faster. 30× less context consumed. Same answer. The LLM learned on its own: "from now on, always search Hippocamp first."
Install
Homebrew (macOS / Linux):
brew install tggo/tap/hippocamp
Direct download — pick your platform from Releases:
# macOS Apple Silicon
curl -L https://github.com/tggo/hippocamp/releases/latest/download/hippocamp_darwin_arm64.tar.gz | tar xz
sudo mv hippocamp /usr/local/bin/
# macOS Intel
curl -L https://github.com/tggo/hippocamp/releases/latest/download/hippocamp_darwin_amd64.tar.gz | tar xz
sudo mv hippocamp /usr/local/bin/
# Linux amd64
curl -L https://github.com/tggo/hippocamp/releases/latest/download/hippocamp_linux_amd64.tar.gz | tar xz
sudo mv hippocamp /usr/local/bin/
Build from source:
git clone https://github.com/tggo/hippocamp
cd hippocamp
make build
Connect to Claude
Add to your MCP config (~/.claude/claude_desktop_config.json or Claude Code settings):
Follow hasTopic, references, partOf links to include related resources (default: false)
explain
Add per-field score breakdown, temporal proximity, and popularity data (default: false)
Searches across rdfs:label, hippo:summary, hippo:alias, hippo:filePath, hippo:signature, hippo:content, hippo:url, hippo:rationale, and subject URIs. Uses field boosting (label matches score highest), word boundary scoring, and score accumulation across predicates.
Use hippo:alias to add synonyms and translations — e.g. Ukrainian labels for English-named resources so search finds them in either language.
With related=true, also returns resources that link TO direct matches via relationship predicates (1-hop graph traversal).
When no results are found, the response includes a hint with the total resource count and suggestions for refining the query.
validate — ontology compliance check
{}
{"scope": "project:house-construction"}
Parameter
Notes
scope
Named graph to validate (omit for all graphs)
Checks:
All rdf:type values are from the hippo: namespace
All typed resources have rdfs:label
All Decisions have hippo:rationale
Fuzzy type matching: typos like hippo:Entiy get a suggestion — "did you mean hippo:Entity? (91% similar)" — with fix commands
Schema migrations: warns if a newer schema version is available with graph action=migrate fix
Returns JSON with valid (bool), warnings (array), fixes (array of suggested commands), and stats.
hippo:source — URI of the agent/tool that produced the resource
Temporal validity:
hippo:validFrom / hippo:validTo — mark facts as time-bounded (invalidated facts stay in graph for history)
The ontology is open-world — extend with your own classes and properties.
Schema migrations
When you upgrade hippocamp, your existing data keeps working (RDF is schema-less). But new features may benefit from enriching old data. The migration system handles this automatically:
Run validate — if a migration is available, you'll see:
"schema update available — run graph action=migrate to apply 1 pending migration(s)"
Run graph action=migrate — applies all pending migrations and reports what changed
The graph's schema version is stored as a triple and checked on each validate
Current migrations:
v2: Adds hippo:provenance "extracted" and hippo:confidence 1.0 to all typed resources that lack them
Migrations are additive — they never delete data. Resources that already have the properties are skipped.
Claude Code integration
Auto-setup
Hippocamp automatically writes hooks and skills to your project's .claude/ directory on first launch. No manual copying needed. On subsequent launches, files are only updated if the binary is newer.
Skill: project-analyze
Ask Claude: "analyze this project and build a knowledge graph". The skill file (.claude/skills/project-analyze.md) is auto-installed and instructs Claude how to scan your project, extract topics, entities, notes, decisions, questions, and sources into the graph.
Hooks
Hooks are auto-installed. You can customize them in .claude/settings.json:
Auto-load on startup from store.default_file (if auto_load: true and file exists)
Manual dump via {"action":"dump","file":"path.trig"}
Auto-save on SIGINT / SIGTERM if there are unsaved changes
Data is stored in TriG format — a superset of Turtle that preserves named graph boundaries.
Named graphs
All tools accept an optional graph parameter (URI string). Omit it to use the default graph.
# Add to named graph
{"action":"add","graph":"http://ex.org/people","subject":"...","predicate":"...","object":"..."}
# Query named graph
{"query":"SELECT ?s WHERE { ?s ?p ?o }","graph":"http://ex.org/people"}
# Create / delete graphs
{"action":"create","name":"http://ex.org/people"}
{"action":"delete","name":"http://ex.org/people"}
Benchmarks
Memory retrieval benchmarks adapted from LongMemEval (ICLR 2025), ConvoMem (Salesforce), and LoCoMo (Snap Research). Measures R@5 (correct result in top 5) using keyword search only, no embeddings.
Benchmark
R@5
Details
MemEval (v2)
100%
extraction 100%, reasoning 100%, temporal 100%
ConvoMem (v2)
100%
user facts 100%, preferences 100%, temporal 100%, implicit 100%