agent-kg 0.7.0

Conversational memory as a knowledge graph

AgentKG — Conversational Memory as a Live, Queryable Knowledge Graph

Author: Eric G. Suchanek, PhD

Flux-Frontiers, Liberty TWP, OH

---

Overview

AgentKG stores every conversation turn, topic, entity, intent, task, and user preference as a node in a persistent knowledge graph (SQLite + LanceDB). Edges encode relationships between turns, sessions, and profile facts. The result is a queryable, prunable, semantically searchable memory that survives context resets and accumulates across projects.

The graph is split into two stores:

• Per-repo conversation graph (.agentkg/) — turns, topics, entities, intents, tasks, summaries
• Global user profile (~/.kgrag/profiles/<person>/) — preferences, expertise, style, commitments, interests; never pruned

Embeddings use all-MiniLM-L6-v2 (384-dim) via sentence-transformers + LanceDB. Structure is treated as ground truth; semantic search is strictly a retrieval accelerant.

---

Features

• Incremental ingest — every turn indexed in real-time; topics, entities, and intents extracted automatically via spaCy + keyword fallback
• Hybrid query — semantic seeding (LanceDB) + structural expansion (graph traversal)
• Global UserProfile tree — preference, expertise, style, commitment, interest, and context nodes accumulated across all repos
• Structured onboarding — four-phase interview populates the profile on first use
• Implicit profile updates — NLP pipeline extracts standing rules from natural language ("always do X", "I prefer Y")
• Context assembly — token-budgeted context block built from the graph for LLM prompt injection
• KG Context Pruning — old turns compressed into Summary nodes when the graph grows large
• Temporal snapshots — point-in-time JSON snapshots for diffing session state
• MCP server — exposes the full query pipeline as structured tools for AI agent integration
• Script-based hooks — three Claude Code hooks (UserPromptSubmit, Stop, PreCompact) deployed as shell scripts via install-hooks

---

Quick Start

# 1. Download the embedding model and create your profile directory
agentkg init --person <you>

# 2. Run the onboarding interview
agentkg onboard --person <you>

# 3. Check your profile
agentkg profile --person <you>

Embeddings are on by default. init pre-warms the model cache so the first ingest does not pause to download.

---

Person ID

--person identifies the global user profile at ~/.kgrag/profiles/<person>/. The default is your OS login name (getpass.getuser()), which is correct automatically on a single-user machine — you rarely need to set it explicitly.

--person only affects profile-scoped commands: init, onboard, profile, viz --profile, wipe --global. Local graph commands (query, assemble, stats, sessions, snapshot, prune, ingest, analyze) are repo-scoped and ignore it.

# Single-user machine: default is correct, no flag needed
agentkg onboard    # writes to ~/.kgrag/profiles/<your-os-username>/
agentkg profile    # reads the same path ← correct

# Multi-user or named profiles: be explicit on profile commands only
agentkg onboard --person alice
agentkg profile --person alice
agentkg query "auth strategy"   # no --person needed here

If agentkg profile returns an empty # UserProfile, check which --person value was used during onboard. The completion message prints the exact path.

---

CLI Reference

All commands accept --repo <path> (default .). --person <id> defaults to your OS username and is only needed for profile-scoped commands (init, onboard, profile, viz --profile, wipe --global).

Command	Description
agentkg init	Download embedding model and create profile directory (run first)
agentkg install-hooks	Deploy hook scripts and wire Claude Code settings.json
agentkg onboard	Run the structured UserProfile onboarding interview
agentkg profile	Show the UserProfile as Markdown
agentkg ingest	Add a turn to the conversation graph
agentkg query	Semantic search over the graph
agentkg assemble	Assemble a token-budgeted context block
agentkg prune	Compress old turns into Summary nodes
agentkg stats	Show graph node/edge counts
agentkg analyze	Print a full Markdown analysis report
agentkg sessions	List all sessions for this repo
agentkg snapshot	Capture a point-in-time snapshot
agentkg mcp	Start the MCP server (stdio transport)

Each command also ships as a dedicated agentkg-<name> script — no poetry run needed:

agentkg-init     --person egs          # profile-scoped
agentkg-onboard  --person egs          # profile-scoped
agentkg-profile  --person egs          # profile-scoped
agentkg-stats    --repo .
agentkg-query    "authentication strategy" --k 8 --repo .
agentkg-assemble "what did we decide about auth?" --budget 4000 --repo .
agentkg-prune    --window 20 --repo .
agentkg-mcp

---

Installation

Requirements: Python ≥ 3.12, < 3.14

Poetry (recommended)

git clone https://github.com/Flux-Frontiers/agent_kg.git
cd agent_kg
poetry install

As a dependency

[tool.poetry.dependencies]
agent-kg = {git = "https://github.com/Flux-Frontiers/agent_kg.git"}

Optional extras

poetry install -E llm    # Anthropic summarizer backend
poetry install -E viz    # Streamlit explorer UI + pyvis graph visualization
poetry install -E local  # httpx for local LLM backends
poetry install -E all    # everything above

spaCy is a required dependency (not an extra) — it is always installed and drives topic and entity extraction. The en_core_web_sm model must be downloaded separately:

python -m spacy download en_core_web_sm

Without the model, extraction falls back to keyword/regex heuristics automatically.

---

Data Layout

<repo-root>/.agentkg/
  graph.sqlite          # nodes + edges (conversation graph)
  lancedb/              # vector embeddings
  snapshots/            # point-in-time JSON snapshots

~/.kgrag/profiles/<person>/
  userprofile.sqlite    # global UserProfile tree (never pruned)

---

Node Kinds

Kind	What it stores
turn	Raw user/assistant message text
topic	N-gram topics extracted from turns
entity	Named entities (people, tools, projects)
intent	Classified intent category per turn
task	Action items extracted from conversation
summary	Pruned turn summaries (after prune pass)
preference	User coding/style preferences (profile)
commitment	Standing rules — "always do X" (profile)
expertise	Domain knowledge areas (profile)
interest	Topics the user cares about (profile)
style	Formatting/docstring/verbosity preferences (profile)
context	Role, machine, projects context (profile)
education	Educational background entries (profile)

---

MCP Server

AgentKG ships a Model Context Protocol (MCP) server that exposes the full query pipeline as structured tools for AI agents.

agentkg-mcp   # stdio transport

Configure in .mcp.json (Claude Code / Kilo Code):

{
  "mcpServers": {
    "agent-kg": {
      "command": "agentkg-mcp"
    }
  }
}

---

Hooks (Auto-Ingest)

AgentKG ships three Claude Code hook scripts that are deployed by the installer:

Hook	Script	What it does
UserPromptSubmit	agent_kg_user_prompt_hook.sh	Ingests each user turn with embeddings
Stop	agent_kg_stop_hook.sh	Ingests assistant turn; runs prune async every 20 exchanges; snapshots
PreCompact	agent_kg_precompact_hook.sh	Runs prune + snapshot synchronously before context compaction — ensures no turns are lost

Install

# Deploy scripts to ~/.agentkg/hooks/ and wire into ~/.claude/settings.json (all repos)
agentkg install-hooks --global

# Or wire into .claude/settings.json for this repo only
agentkg install-hooks --claude

# Force-overwrite existing hooks
agentkg install-hooks --global --force

The installer:

1. Copies the three .sh scripts from the package into ~/.agentkg/hooks/ (executable)
2. Merges UserPromptSubmit, Stop, and PreCompact entries into the target settings.json

The scripts are portable — they use git rev-parse --show-toplevel to locate the repo and only fire when a .agentkg/ directory is present.

Hook state and logs

~/.agentkg/hook_state/
  hook.log                          # timestamped log of all hook activity
  <session_id>_last_consolidate     # exchange counter for periodic prune

---

Project Structure

agent_kg/
├── README.md
├── pyproject.toml
├── hooks/                            # reference copies of hook scripts
│   ├── agent_kg_user_prompt_hook.sh
│   ├── agent_kg_stop_hook.sh
│   └── agent_kg_precompact_hook.sh
├── scripts/
│   └── generate_wiki.py              # GitHub wiki generator
├── src/
│   └── agent_kg/
│       ├── __init__.py
│       ├── graph.py                  # AgentKG orchestrator
│       ├── store.py                  # SQLite + LanceDB storage
│       ├── index.py                  # LanceDB semantic indexing
│       ├── ingest.py                 # Phase 1 incremental turn ingest
│       ├── user_profile.py           # Global UserProfile tree
│       ├── onboard.py                # Structured onboarding interview
│       ├── session.py                # Session lifecycle
│       ├── query.py                  # Hybrid semantic + graph query
│       ├── assemble.py               # Token-budgeted context assembly
│       ├── prune.py                  # KG Context Pruning
│       ├── consolidate.py            # Deferred embedding consolidation
│       ├── summarize.py              # LLM-backed summarization
│       ├── snapshots.py              # Point-in-time snapshot capture
│       ├── schema.py                 # Node/Edge dataclasses
│       ├── kg.py                     # High-level KG facade
│       ├── app.py                    # Streamlit explorer UI
│       ├── viz.py                    # Visualization helpers (Rich + pyvis)
│       ├── hooks/                    # bundled hook scripts (deployed by install-hooks)
│       │   ├── agent_kg_user_prompt_hook.sh
│       │   ├── agent_kg_stop_hook.sh
│       │   └── agent_kg_precompact_hook.sh
│       ├── cli/
│       │   ├── main.py               # Click CLI entry points
│       │   └── __init__.py
│       ├── mcp/
│       │   └── server.py             # MCP server
│       └── nlp/                      # NLP pipeline (spaCy + regex fallback)
│           ├── entities.py
│           ├── intent.py
│           ├── preferences.py
│           └── topics.py
└── tests/

---

License

Elastic License 2.0 — see LICENSE.

Free to use, modify, and distribute. You may not offer the software as a hosted or managed service to third parties. Commercial use internally is permitted.