hippocamp

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.

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):

{
  "mcpServers": {
    "hippocamp": {
      "command": "hippocamp",
      "args": ["--config", "/path/to/config.yaml"]
    }
  }
}

Tools

triple — add / remove / list triples

add:    {"action":"add","subject":"http://ex.org/Alice","predicate":"http://ex.org/name","object":"Alice","object_type":"literal"}
remove: {"action":"remove","subject":"http://ex.org/Alice","predicate":"http://ex.org/name","object":"http://ex.org/Bob"}
list:   {"action":"list","subject":"http://ex.org/Alice"}

Parameter	Values	Notes
action	add | remove | list	required
graph	URI string	optional, omit for default graph
subject	URI	required for add/remove, wildcard if empty for list
predicate	URI	required for add/remove, wildcard if empty for list
object	string	required for add/remove, wildcard if empty for list
object_type	uri | literal | bnode	default: uri
lang	e.g. en	language tag for literals
datatype	XSD URI	e.g. http://www.w3.org/2001/XMLSchema#integer

---

sparql — SELECT / ASK / UPDATE

SELECT: {"query": "SELECT ?s ?p ?o WHERE { ?s ?p ?o } LIMIT 10"}
ASK:    {"query": "ASK { <http://ex.org/Alice> a <http://ex.org/Person> }"}
UPDATE: {"query": "INSERT DATA { <http://ex.org/Alice> <http://ex.org/name> \"Alice\" }"}
Named:  {"query": "SELECT ?name WHERE { ?s <http://ex.org/name> ?name }", "graph": "http://ex.org/g1"}

Returns JSON bindings for SELECT, "true"/"false" for ASK, "ok" for updates.

---

graph — graph lifecycle, persistence, prefixes

Graph management:

{"action":"list"}
{"action":"create","name":"http://example.org/myGraph"}
{"action":"delete","name":"http://example.org/myGraph"}
{"action":"stats","name":"http://example.org/myGraph"}
{"action":"clear","name":"http://example.org/myGraph"}

Persistence:

{"action":"dump","file":"./backup.trig"}
{"action":"load","file":"./backup.trig"}

Overview and migration:

{"action":"summary"}   — compact graph overview (~500 tokens): type counts, top entities, topics, recent decisions
{"action":"migrate"}   — apply pending schema migrations (e.g. add provenance defaults to old data)

Namespace prefixes:

{"action":"prefix_add","prefix":"ex","uri":"http://example.org/"}
{"action":"prefix_list"}
{"action":"prefix_remove","prefix":"ex"}

---

search — semantic keyword search

{"query": "authentication"}
{"query": "Store", "type": "https://hippocamp.dev/ontology#Struct"}
{"query": "middleware", "scope": "project:myapp", "limit": 5}

Parameter	Notes
query	Search keywords (required, case-insensitive)
type	Filter by rdf:type URI
scope	Named graph to search in (omit for all graphs)
limit	Max results (default: 20)
related	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.

---

analyze — graph structure analysis

{"action":"god_nodes","limit":5}
{"action":"components"}
{"action":"surprising"}
{"action":"consolidate","limit":10}
{"action":"export_html"}

Action	What it does
god_nodes	Most-connected resources by degree, excluding Topic/Tag/Project hubs
components	Connected components (clusters) via BFS
surprising	Cross-topic or cross-graph edges (bridges between clusters)
consolidate	Find resources with missing summaries/topics, suggest enrichments with graph context
export_html	Return URL of the live visualization server (auto-started on port 39322)

Parameter	Notes
action	required
scope	Named graph to analyze (omit for all graphs)
limit	Max results for god_nodes (default: 10) and consolidate (default: 20)

---

CLI query mode

For scripts and hooks, run a one-shot search without starting the MCP server:

hippocamp --config config.yaml --query "search terms"
hippocamp --config config.yaml --query "auth" --type "https://hippocamp.dev/ontology#Function" --limit 5

Loads the persisted graph, runs the search, prints pretty-printed JSON, and exits.

Ontology

Hippocamp includes a lightweight RDF ontology (ontology/hippo.ttl) with two layers:

Base layer (any domain):

• hippo:Topic — subject areas, themes
• hippo:Entity — people, orgs, products, APIs, datasets
• hippo:Note — free-form observations
• hippo:Source — articles, papers, URLs, books
• hippo:Decision — recorded decisions with rationale
• hippo:Question — open questions
• hippo:Tag — lightweight labels

Code layer (software projects):

• hippo:Project, hippo:Module, hippo:File
• hippo:Function, hippo:Struct, hippo:Interface, hippo:Class
• hippo:Dependency, hippo:Concept

Provenance and quality tracking:

• hippo:confidence — float 0.0-1.0 (how certain is this fact?)
• hippo:provenance — "extracted" / "inferred" / "ambiguous"
• 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:

1. Run validate — if a migration is available, you'll see:

   "schema update available — run graph action=migrate to apply 1 pending migration(s)"
   

2. Run graph action=migrate — applies all pending migrations and reports what changed
3. 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:

{
  "hooks": {
    "UserPromptSubmit": [
      { "command": ".claude/hooks/hippocamp-pre-query.sh \"$PROMPT\"" }
    ],
    "PostToolUse": [
      {
        "command": ".claude/hooks/hippocamp-post-edit.sh \"$TOOL_NAME\" \"$FILE_PATH\"",
        "matcher": "Edit|Write"
      }
    ]
  }
}

• Pre-query hook: Before each prompt, queries the graph for relevant context
• Post-edit hook: After file edits, tracks which files need re-indexing

Configuration

# config.yaml
store:
  default_file: "./data/default.trig"  # loaded on startup, saved on SIGINT/SIGTERM
  auto_load: true
  format: "trig"

prefixes:
  ex: "http://example.org/"
  schema: "http://schema.org/"

ENV overrides:

Variable	Default
HIPPOCAMP_STORE_DEFAULT_FILE	./data/default.trig
HIPPOCAMP_STORE_AUTO_LOAD	true
HIPPOCAMP_STORE_FORMAT	trig

Persistence

• 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%
LoCoMo (v2)	88.9%	single-hop 100%, multi-hop 100%, temporal 100%, adversarial 0%

Known weakness: adversarial/abstention queries (keyword search can't distinguish "no relevant result" from partial keyword matches).

Run benchmarks:

go test ./internal/tools/ -run TestMemoryBenchmark -v

Visualization

Hippocamp starts an HTTP server on localhost:39322 alongside the MCP server. Open it in a browser for an interactive graph visualization with:

• Nodes colored by type, sized by degree
• Search and type filter
• Click-to-inspect with full property list and edge details

Development

make test      # run all tests
make build     # build binary
make lint      # go vet
make run       # build + run with config.yaml