Project Knowledge Graph

Cross-project knowledge graph powered by local FalkorDB. Indexes all project artifacts (recaps, plans, project memory files, architecture docs, skills) across multiple projects and lets you query by concept. "What did we learn about FalkorDB replication?" returns the matching skill, relevant recaps, and architecture context — regardless of which project they came from.

A local FalkorDB-backed semantic index over all project artifacts. Query by concept across 8 projects: "show me everything about FalkorDB replication" returns the matching skill, the recaps that mention it, and the architecture docs — ranked by TF-IDF relevance with Cypher CONTAINS primary filter.

Architecture

text

knowledge index
    │
    ▼
Scans project directories for .md artifacts
    ├── docs/recaps/, docs/daily-recaps/
    ├── docs/plans/
    ├── project memory file
    ├── docs/architecture/, docs/features/, docs/operations/, docs/pipeline/
    ├── TECHNICAL-DOCUMENTATION.md, FUNCTIONAL-SPECIFICATIONS.md
    └── **/SKILL.md (Hermes skills)
    │
    ▼
Chunks by heading-2 boundaries + paragraphs (max 2000 chars)
    │
    ▼
MERGE into FalkorDB graph as :Chunk nodes
    │
    ▼
knowledge query "concept"
    │
    ▼
Stage 1: Cypher CONTAINS (fast primary filter)
Stage 2: TF-IDF re-ranking
    │
    ▼
Ranked results with project, file, heading, snippet

Setup (one-time)

1. Start the FalkorDB container

bash

docker run -d \
  --restart=unless-stopped \
  -p 127.0.0.1:16379:6379 \
  -v knowledge-graph-data:/data \
  --name knowledge-graph \
  falkordb/falkordb:latest

🔒 Security: The port is bound to 127.0.0.1 (localhost only) — FalkorDB is not reachable from other machines on your network. No authentication is configured because the service is only accessible to local processes.

📦 Docker image note: falkordb/falkordb:latest is a mutable tag. To pin by digest (recommended for production), replace the tag with the current SHA256 after pulling:

bash

docker pull falkordb/falkordb:latest
DIGEST=$(docker inspect --format='{{index .RepoDigests 0}}' falkordb/falkordb:latest)
echo "Replace :latest with @${DIGEST#*@}"

Then use falkordb/falkordb@sha256:... in the docker run command. Check for updates intentionally rather than relying on automatic pulls.

Auto-starts on Docker daemon start via --restart=unless-stopped. Data persists in the Docker volume.

ℹ️ Persistence note: The container uses --restart=unless-stopped to survive Docker restarts but won't auto-restart after a manual docker stop. To disable persistence entirely, omit --restart and start the container manually when needed.

2. Install the Python dependency

bash

pip install falkordb==1.6.1

3. Verify

bash

python3 ~/.hermes/scripts/project-knowledge-index.py doctor

Should show all 8 project roots with document counts and a healthy FalkorDB connection.

Usage

Index all projects

bash

# First index (~60s) — incremental thereafter (<1s)
python3 ~/.hermes/scripts/project-knowledge-index.py index

# Preview without writing
python3 ~/.hermes/scripts/project-knowledge-index.py index --dry-run

⚠️ Before first real run: Use --dry-run to preview which files would be indexed. This is especially important when configuring PROJECT_ROOTS for the first time — it shows you exactly which files the scraper will read before any data is written to FalkorDB.

bash

python3 ~/.hermes/scripts/project-knowledge-index.py index --dry-run

Query by concept

bash

# Basic search
python3 ~/.hermes/scripts/project-knowledge-index.py query "FalkorDB replication"

# Filter by project
python3 ~/.hermes/scripts/project-knowledge-index.py query "batch writes" --project CI

# Filter by document type (recap, plan, skill, claude, architecture)
python3 ~/.hermes/scripts/project-knowledge-index.py query "soft delete" --type skill

Stats and health

bash

python3 ~/.hermes/scripts/project-knowledge-index.py stats
python3 ~/.hermes/scripts/project-knowledge-index.py doctor

Query Algorithm

Two-stage ranking:

1. Cypher CONTAINS — finds all chunks whose text contains the search terms. Fast, exact-match primary filter. Supports filtering by --project and --type at this stage.
2. TF-IDF re-ranking — computes term frequency × inverse document frequency for each candidate against the query tokens. Scores are relative within the result set (best match = highest score).

This gives substantially better results than pure CONTAINS without needing LLM embeddings or API calls.

How to use this in a session

When working on a problem you've seen before:

bash

# How did we handle PostgreSQL migrations?
python3 ~/.hermes/scripts/project-knowledge-index.py query "PostgreSQL migration"

# What patterns exist for batch operations across projects?
python3 ~/.hermes/scripts/project-knowledge-index.py query "batch" --type skill

# Did we ever document Ollama rate limiting?
python3 ~/.hermes/scripts/project-knowledge-index.py query "Ollama rate limit"

# What did we learn about soft deletes?
python3 ~/.hermes/scripts/project-knowledge-index.py query "soft delete" --project MyApp

What's indexed

Project	Typical docs	Sample content
Any project with a project memory file	Recaps, plans, memory file	Session recaps, implementation plans, methodology rules
Any project with docs/	Architecture, features, operations, pipeline	Topical docs about every component
Any project with TECHNICAL/FUNCTIONAL specs	Full technical specs	Architecture decisions, schema docs
Hermes skills (~/.hermes/skills/)	All SKILL.md files	Every installed skill — procedures, tips, workflows
Hermes Agent repo	Built-in skills, docs, plans	Stock skills, plugin docs

Why FalkorDB over SQLite

Dimension	SQLite (hash)	FalkorDB (graph — chosen)
Primary filter	Simhash distance (fuzzy)	Cypher CONTAINS (exact match)
Cross-project query	Manual JOINs	MATCH (c:Chunk {project:'CI'}) RETURN c
Data model flexibility	Fixed SQL schema	Add node types, edges on the fly
Infrastructure	File on disk	Single Docker container (localhost-only, 200MB idle)
Query speed	Sub-second	Sub-second

The Docker container is a one-time setup. After that, it starts on boot automatically (--restart=always) and uses <200MB RAM when idle.

Configuring Your Projects

Edit PROJECT_ROOTS in ~/.hermes/scripts/project-knowledge-index.py to point at your own project directories:

python

PROJECT_ROOTS = {
    "MyApp": os.path.expanduser("~/Projects/myapp"),
    "Website": os.path.expanduser("~/Desktop/website"),
    "Skills": os.path.expanduser("~/.hermes/skills"),
    "Hermes": os.path.expanduser("~/.hermes/hermes-agent"),
}

The Custom-Skills and Hermes-Agent entries are optional but recommended — they index every Hermes skill you have (both installed and stock) so you can search across all known patterns.

Pro tip: Add your projects in priority order. The find_documents() function respects your glob patterns — it specifically scopes to docs/recaps/*.md, docs/plans/*.md, project memory files, docs/architecture/*.md, docs/features/*.md, SKILL.md, etc. It does NOT crawl node_modules/.

Pitfalls

1. CONTAINS is case-sensitive

"FalkorDB" matches, "falkordb" does not. Use the casing you expect in documents. Document text preserves original casing from markdown files.

2. Short queries produce noisy results

"DB" or "pipeline" match too broadly. Mitigation: add --project or --type filters, or use more specific terms (e.g., "FalkorDB replication" not just "FalkorDB").

3. TF-IDF scores are relative within result set

A score of 0.06 on a 10-result query doesn't mean "6% match" — it means "the top result by far." Use scores to rank within a single query, not to compare across queries.

4. FalkorDB container must be running

bash

docker start knowledge-graph  # if stopped

5. Massive node_modules directories hang rglob-based scans

The initial version used Path.rglob("*.md") to count files in doctor. Projects with large node_modules directories (Beacon-v2 had 775+ entries in one subdirectory with symlink loops) caused 30s+ hangs. Fix: never rglob a project root. Use targeted root.glob(pattern) with specific paths like docs/recaps/*.md, docs/plans/*.md. The find_documents function already does this correctly — the original doctor function was the only offender.

6. First index is slow (~60s)

Subsequent runs are <1s (only new/changed files are re-indexed). The indexer uses content hash + mtime for change detection.

6. The skill indexes itself

Since the project-knowledge-graph SKILL.md is in ~/.hermes/skills/, it gets indexed. Query results may show matches from this skill's documentation — usually harmless.

Data Retention & Purge

Indexed content persists in the Docker volume until explicitly removed. This means cross-project knowledge is available across sessions without re-indexing, but stale or sensitive content remains searchable until purged.

Important: the indexer uses MERGE (upsert), not sync. When a source file is modified, its chunks are re-indexed on the next knowledge index run. However, if a file is deleted or a chunk is removed, the old chunks remain in FalkorDB until the volume is purged. The indexer does not track deletions — it only adds and updates.

If you need to guarantee no stale content remains after removing source files:

bash

# Full purge and rebuild
docker stop knowledge-graph && docker rm knowledge-graph && docker volume rm knowledge-graph-data
docker run -d --restart=unless-stopped -p 127.0.0.1:16379:6379 -v knowledge-graph-data:/data --name knowledge-graph falkordb/falkordb:latest
python3 ~/.hermes/scripts/project-knowledge-index.py index

Or use the delete command for targeted removal without stopping the container:

bash

# Delete all chunks for a specific project
python3 ~/.hermes/scripts/project-knowledge-index.py delete --project CI

# Delete all chunks of a specific type across all projects
python3 ~/.hermes/scripts/project-knowledge-index.py delete --type recap

# Delete all chunks (rebuild from scratch)
python3 ~/.hermes/scripts/project-knowledge-index.py delete --all
python3 ~/.hermes/scripts/project-knowledge-index.py index

Stop the service (keep data)

bash

docker stop knowledge-graph

To re-enable: docker start knowledge-graph

Clear all indexed data (stop + purge)

bash

docker stop knowledge-graph && docker rm knowledge-graph && docker volume rm knowledge-graph-data

After purging, re-run knowledge index (first run will be ~60s to rebuild the index).

Selective re-indexing (update specific projects)

The indexer uses content hashing — unchanged documents are skipped automatically. To force a full re-index of all projects (e.g. after changing PROJECT_ROOTS):

bash

# Option A: Clear and re-index
docker stop knowledge-graph && docker rm knowledge-graph && docker volume rm knowledge-graph-data
docker run -d --restart=unless-stopped -p 127.0.0.1:16379:6379 -v knowledge-graph-data:/data --name knowledge-graph falkordb/falkordb:latest
python3 ~/.hermes/scripts/project-knowledge-index.py index

# Option B: Just re-index (updates changed files, retains unchanged)
python3 ~/.hermes/scripts/project-knowledge-index.py index

CLI Reference

text

usage: knowledge index|query|stats|doctor|delete

Commands:
  index       Scan and index all project documents
              --dry-run  Preview without indexing

  query       Search the knowledge graph by concept
              terms            Search terms (required)
              --project, -p    Filter by project name
              --type, -t       Filter by doc type (recap/plan/skill/claude/architecture)
              --limit, -l      Max results (default: 10)

  stats       Show corpus statistics (chunks per project/type)

  doctor      Check environment and connectivity

  delete      Delete chunks from the graph
              --project, -p    Delete all chunks for a project
              --type, -t       Delete all chunks of a type
              --all            Delete ALL chunks (equivalent to docker volume purge)

Indexed Document Types

Type	What's included
recap	docs/recaps/*.md, docs/daily-recaps/*.md
plan	docs/plans/*.md
claude	Project memory files (e.g. PROJECT.md)
architecture	docs/architecture/*.md, docs/features/*.md, docs/operations/*.md, docs/pipeline/*.md, TECHNICAL-DOCUMENTATION.md, FUNCTIONAL-SPECIFICATIONS.md
skill	**/SKILL.md anywhere in project or ~/.hermes/skills/

Auto-projection (via session-wrapup)

New knowledge is projected into the graph automatically at session end. The session-wrapup skill runs knowledge index as one of its steps, which picks up:

• New recaps — every session recap gets indexed automatically
• Updated plans — plan status changes and criteria updates get indexed
• Modified project memory files — any "Today's state" updates get indexed
• New/changed skills — any skill edited or created gets indexed

Additionally, wrapup queries the graph with key terms from the session's work and surfaces cross-project connections as "Did you know?" findings in the wrap-up report.

No manual steps needed. The knowledge graph stays current as a side effect of the normal session lifecycle.

• knowledge doctor shows all project roots and healthy FalkorDB connection
• knowledge index completes and reports chunk count
• knowledge query "some concept" returns ranked results with project/file/snippet
• knowledge query --project CI --type plan filters correctly
• Incremental: running knowledge index twice skips unchanged files
• Container auto-starts on Docker daemon restart
• Cross-project: query returns results from multiple projects