waggle-mcp 0.0.1

MCP server that gives LLMs persistent graph-structured memory

waggle-mcp

Persistent memory that remembers decisions, reasons, and contradictions across sessions.
Your AI forgets everything when the context window closes. Waggle gives it a graph-backed brain that persists.

Not a code indexer. A conversational memory engine — it stores what you decided, why you decided it, and what changed, so the next session picks up where the last one left off.

---

Core

This repository is the public Waggle product repo: Apache-2.0 licensed, available on GitHub and PyPI, and focused on the local-first memory engine.

---

Quick Start

# Install globally (no venv needed)
pipx install waggle-mcp

# One-line setup — detects your MCP clients and writes config
waggle-mcp setup --yes

# Verify everything is healthy
waggle-mcp doctor

(No pipx? Run brew install pipx && pipx ensurepath first.)

setup --yes detects Claude Code, Codex, Cursor, Gemini CLI, and Antigravity, writes the MCP config, and installs automatic memory hooks where supported. Restart your client and you're live.

Windows users: Run all commands with python -X utf8 or set PYTHONUTF8=1 to avoid UnicodeEncodeError from emoji in log output.

Enterprise Evaluation

For self-hosted production review and security posture:

• Production deployment guide
• Security model
• Hardening checklist
• Reference

---

60-Second Demo

No MCP client needed. Run this from a fresh install:

waggle-mcp demo

This imports a pre-loaded example graph and runs 4 scripted queries locally — no API key, no network, no client required. Add --with-embeddings to use the real sentence-transformers model for higher-fidelity retrieval (requires ~420 MB download on first run).

---

Why Waggle

waggle-mcp is a local-first memory layer for MCP-compatible AI clients, built on a persistent knowledge graph.

The core difference from flat note storage or chunked RAG is the graph structure. Waggle doesn't just store facts — it stores the relationships between them: this decision depends on that constraint, this preference contradicts that earlier one, this requirement was updated three sessions ago. When you query, you get a subgraph with the reasoning chain attached, not just the matching text.

Without Waggle	With Waggle
Paste context into every session	Compact subgraph retrieved at query time
Session-local memory only	Persistent memory across all sessions
Flat notes, no structure	Typed nodes and edges: decisions, reasons, contradictions
"What changed?" requires replaying logs	Temporal queries, diffs, and conflict resolution are first-class
Contradictions silently overwrite history	Both positions preserved, contradiction edge explicit

What Is In Core Today

Waggle Core is the open-source local memory foundation:

• SQLite-backed graph memory
• MCP server integration
• CLI setup and doctor flows
• local embeddings or deterministic fallback
• graph querying, observation, and context priming
• import/export and graph inspection utilities

Product Scope

This public repo is the product-facing Waggle surface:

• MCP server and tool surface
• local-first graph memory
• automatic memory hooks and orchestration
• .abhi export, import, diff, merge, and checkpoint handoff
• Graph Studio and admin tooling

Research artifacts, benchmark harnesses, evaluation reports, and paper material now live in the private waggle-pro repo.

---

Architecture

MCP Client (Claude / Codex / Gemini CLI / Cursor / Antigravity / ChatGPT)
    ↓
waggle.server  — MCP tool surface
    ↓
RecursiveContextController  — RLM-inspired context assembly (build_context)
    ↓
Graph Engine  — MemoryGraph (SQLite) or Neo4jMemoryGraph
    ↓
Embeddings  — sentence-transformers (local) or deterministic fallback

---

Recursive Context Assembly

Waggle stores memory outside the model context window. Instead of pasting long context into every prompt, agents call build_context to get a compact, high-signal context pack assembled from the graph.

Inspired by Recursive Language Models — the idea of externalising long context into an environment and interacting with it through decomposition and targeted retrieval.

How it works:

1. Decompose — the query is split into targeted subqueries (decisions, constraints, implementation details, unfinished work, conflicts)
2. Retrieve — each subquery runs against graph, hybrid, and verbatim transcript retrieval
3. Expand — the graph is traversed around top nodes via typed edges (updates, contradicts, depends_on, derived_from)
4. Resolve — update chains and contradictions are detected; superseded nodes are flagged
5. Deduplicate & rank — overlapping hits are merged; high-signal node types (decisions, preferences) are boosted
6. Compress — everything is packed into a structured context brief under a configurable token budget

Example MCP call:

{
  "tool": "build_context",
  "arguments": {
    "query": "Continue implementing Waggle from where we left off",
    "project": "waggle-mcp",
    "token_budget": 1000,
    "depth": 2
  }
}

Example output:

### Waggle Recursive Context Pack
Task: Continue implementing Waggle from where we left off

Current relevant decisions:
- [decision] Use SQLite for local storage: We chose SQLite with WAL mode for local-first deployments.
- [decision] Hybrid retrieval default: Hybrid (vector + BM25 + graph) is the default retrieval mode.

Active constraints:
- [preference] No external LLM APIs required: All retrieval must work fully local.

Important implementation context:
- [fact] RecursiveContextController added: New module waggle/recursive_context.py implements build_context.

Conflicts or superseded context:
- Possible conflict: 'Use Flask' contradicts 'Use FastAPI'

Config env vars:

Variable	Default	Description
WAGGLE_RECURSIVE_CONTEXT_ENABLED	true	Enable/disable the feature
WAGGLE_RECURSIVE_CONTEXT_DEFAULT_BUDGET	1200	Default token budget
WAGGLE_RECURSIVE_CONTEXT_MAX_SUBQUERIES	6	Max decomposed subqueries
WAGGLE_RECURSIVE_CONTEXT_DEFAULT_DEPTH	2	Graph expansion depth
WAGGLE_RECURSIVE_CONTEXT_INCLUDE_EVIDENCE	true	Include transcript evidence

Tool aliases: recursive_context, assemble_context, rlm_context all resolve to build_context.

---

How It Works

User  → Agent → observe_conversation(...)  → Graph stores typed nodes + edges
User  → Agent → query_graph("database")   → Subgraph returned → Agent answers with linked rationale

Session 1

User:  Let's use PostgreSQL. MySQL replication has been painful.
Agent: [calls observe_conversation()]
       → stores decision node: "Chose PostgreSQL over MySQL"
       → stores reason node:   "MySQL replication painful"
       → links them with a depends_on edge

Session 2 (fresh context window, no history)

User:  What did we decide about the database?
Agent: [calls query_graph("database decision")]
       → retrieves the decision node + linked reason from Session 1
       "You decided on PostgreSQL. The reason recorded was that MySQL replication had been painful."

Session 3

User:  Actually, let's reconsider — the team is more familiar with MySQL.
Agent: [calls store_node() + store_edge(new_node → old_node, "contradicts")]
       → both positions are preserved, and the contradiction is explicit

---

Setting Up as an MCP Server

One-time install: pipx install waggle-mcp — no API key, no cloud account, no Docker required for local use.

Shared JSON config for clients that accept mcpServers JSON:

{
  "mcpServers": {
    "waggle": {
      "command": "waggle-mcp",
      "args": ["serve"],
      "env": {
        "WAGGLE_TRANSPORT": "stdio",
        "WAGGLE_BACKEND": "sqlite",
        "WAGGLE_DB_PATH": "~/.waggle/memory.db",
        "WAGGLE_DEFAULT_TENANT_ID": "local-default",
        "WAGGLE_MODEL": "all-MiniLM-L6-v2",
        "WAGGLE_STARTUP_MODE": "normal"
      }
    }
  }
}

First run takes ~30 s — all-MiniLM-L6-v2 (~420 MB) downloads on first use. To skip the download: set "WAGGLE_MODEL": "deterministic" (offline-safe, instant start, slightly lower retrieval quality).

Claude Desktop

Config file location:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add the mcpServers block above.

Claude Code

claude mcp add waggle \
  --env WAGGLE_TRANSPORT=stdio \
  --env WAGGLE_BACKEND=sqlite \
  --env WAGGLE_DB_PATH=~/.waggle/memory.db \
  --env WAGGLE_DEFAULT_TENANT_ID=local-default \
  --env WAGGLE_MODEL=all-MiniLM-L6-v2 \
  -- waggle-mcp serve

Claude Code also supports automatic memory hooks — see the Hooks section below.

Codex

Add to ~/.codex/config.toml:

[mcp_servers.waggle]
command = "waggle-mcp"
args    = ["serve"]
env     = {
  WAGGLE_TRANSPORT         = "stdio",
  WAGGLE_BACKEND           = "sqlite",
  WAGGLE_DB_PATH           = "~/.waggle/memory.db",
  WAGGLE_DEFAULT_TENANT_ID = "local-default",
  WAGGLE_MODEL             = "all-MiniLM-L6-v2"
}

waggle-mcp setup --yes also writes a managed memory block into AGENTS.md in the current workspace so automatic memory is enabled by default for that repo.

Gemini CLI

gemini mcp add waggle \
  -e WAGGLE_TRANSPORT=stdio \
  -e WAGGLE_BACKEND=sqlite \
  -e WAGGLE_DB_PATH=~/.waggle/memory.db \
  -e WAGGLE_DEFAULT_TENANT_ID=local-default \
  -e WAGGLE_MODEL=all-MiniLM-L6-v2 \
  waggle-mcp serve

After restarting, run /mcp to confirm Waggle is connected.

Cursor

Cursor Settings → Features → MCP Servers → + Add

• Command: waggle-mcp
• Args: serve
• Env vars: same keys as the JSON block above.

Antigravity

The AI agent reads ~/.gemini/antigravity/mcp_config.json (macOS/Linux) or %USERPROFILE%\.gemini\antigravity\mcp_config.json (Windows). Add the waggle block there. The VS Code extension panel reads a different file — adding waggle there will NOT make it available to the AI agent.

Run waggle-mcp doctor to see exactly which config files exist and which ones have a waggle entry.

ChatGPT

ChatGPT custom MCP connectors require a remote HTTPS server. Deploy Waggle in HTTP mode with the Neo4j backend, expose /mcp over HTTPS, then add that URL as a custom connector in ChatGPT (Settings → Connectors → Advanced → Developer mode).

WAGGLE_TRANSPORT=http \
WAGGLE_BACKEND=neo4j \
WAGGLE_DEFAULT_TENANT_ID=workspace-default \
WAGGLE_NEO4J_URI=bolt://localhost:7687 \
WAGGLE_NEO4J_USERNAME=neo4j \
WAGGLE_NEO4J_PASSWORD=change-me \
waggle-mcp serve

Do not expose Waggle publicly without authentication.

waggle-mcp not on PATH?

pipx ensurepath   # then restart your terminal

---

Automatic Memory — Prompt Rules

Registering Waggle as an MCP server only makes the tools available. For the agent to call them automatically, add this instruction block to your client's prompt, rules, or project instructions:

Use Waggle automatically for conversational memory.

At the start of a new session, if project, agent, or session scope is known, call prime_context.

Before answering questions that may depend on prior decisions, preferences, constraints, project state,
or earlier conversation context, call query_graph with the narrowest relevant scope.

After completed turns that contain durable information such as decisions, preferences, constraints,
requirements, user corrections, project facts, or meaningful task outcomes, call observe_conversation
automatically.

Waggle should remember relevant context automatically. If memory appears empty, the session is likely
missing the automatic memory policy or the runtime hooks that call build_context before answers and
on_assistant_turn after answers.

Do not ask the user to trigger Waggle manually. Use it in the background when relevant.

Use the same stable project value for the same codebase across sessions, or recall will fragment.

---

Automatic Memory Hooks (Claude Code)

For Claude Code, waggle-mcp setup --yes installs three hook scripts that capture memory deterministically — no prompt rules needed:

Hook script	Claude Code event	What it does
pre_response.py	UserPromptSubmit	Tries scoped DB recall first; if the scope is cold and a session checkpoint exists, imports the .abhi checkpoint and retries before Claude responds
post_response.py	Stop	Applies Waggle's durable-ingest policy and only calls observe_conversation for turns worth remembering
pre_compact.py	PreCompact	Calls ingest_transcript_handoff to preserve durable info before context compression, emit a session .abhi checkpoint, and refresh the checkpoint manifest

Each hook always exits 0 (a Waggle bug never blocks your session) and has a 5-second timeout. post_response.py scans turn text for likely secrets before storing, skips low-value chatter, and only ingests durable turns.

# Install hooks (included in setup --yes)
waggle-mcp setup --yes

# Skip hook installation
waggle-mcp setup --yes --no-hooks

# Remove hooks
waggle-mcp uninstall-hooks

---

Verify It Works

After restarting your client, ask the agent:

"Store a note: we're using PostgreSQL for this project."

Then open a fresh session and ask:

"What database are we using?"

Expected:

You're using PostgreSQL for this project.

---

MCP Tool Reference

The full tool surface is large (~40 tools). In practice, an agent in normal use only needs the six core tools. Everything else is for human-driven inspection, graph management, and export workflows.

Core tools — what the agent calls automatically

These are the tools your prompt rules or hooks should wire up. An agent that only knows these six will handle the vast majority of memory tasks correctly.

Tool	When the agent calls it
observe_conversation	After any turn containing a decision, preference, constraint, correction, or project fact. Persists the verbatim turn first, then extracts graph nodes. Returns turn_id, verbatim_stored, nodes_extracted, edges_inferred.
query_graph	Before answering questions that may depend on prior context. Hybrid retrieval (graph + verbatim transcript) by default. Supports as_of for point-in-time queries.
prime_context	At the start of a new session to hydrate context from the most relevant scoped memories.
graph_diff	When the user asks what changed recently.
store_node + store_edge	When the agent needs to store a single atomic fact or explicitly link two nodes. Prefer observe_conversation for conversational turns — use these for structured, deliberate writes.

observe_conversation and decompose_and_store create edges automatically. If you only call store_node, you get isolated facts with no traversal value.

Extended retrieval — agent-callable, situational

Tool	Description
aggregate_graph	Broad filtered subgraph for map-reduce tasks. Use when you want a large scoped slice rather than high-precision top-K. Supports node_types, tags, as_of, include_invalidated.
get_related	Fetch the neighborhood around a specific node by ID.
get_node_history	Inspect a node's evidence, validity window, and connected context.
get_topics	Topic clusters via community detection across the full tenant graph.
timeline	Chronological view of memory changes for a node or query.
list_conflicts	List unresolved contradiction and update edges.
resolve_conflict	Mark a conflict resolved. Pass winner (node ID) to supersede the losing node — sets its valid_to to now.

Operator / human tools — not for routine agent use

These are for you as the operator: graph health, deduplication, export, and migration. Exposing all of these to an agent in normal use adds noise without benefit.

Graph management

Tool	Description
update_node	Update an existing node's content, label, or tags.
delete_node	Delete a node and all its edges.
decompose_and_store	Break long content into atomic nodes and infer edges automatically.
dedup_candidates	Return near-duplicate node pairs above a similarity threshold for human review.
canonicalize_node	Merge multiple nodes into one canonical node. Repoints all edges, collects aliases. Idempotent.
edge_quality_report	Audit edge quality — counts, average confidence per type, top/bottom confidence edges.
debug_retrieval	Diagnose retrieval ranking for a query — embedding scores, window routing, tiered vs flat comparison.

Context windows

Tool	Description
list_context_scopes	Known agent, project, and session scope values.
list_context_windows	Chat/session-level memory containers with status and node counts.
get_context_window	Inspect one context window and its nodes.
close_context_window	Close a session window and derive cross-window edges.
window_graph_viz	Export the context-window graph as an interactive HTML visualization.
export_graph_html	Export the memory graph as an interactive HTML visualization.
get_stats	Node/edge counts, type breakdowns, and recent highly-connected nodes.

Memory files — git-vocabulary interface

Waggle uses a git-inspired vocabulary for portable memory snapshots:

Tool	Git analogy	Description
commit	git commit	Snapshot the graph to a .abhi file. Formats: abhi (default), backup (raw JSON), bundle (Markdown/JSON context pack).
pull	git pull	Load a .abhi or backup file into the current graph. Runs integrity verification before merging.
diff	git diff	Compare two .abhi files — added/removed/updated nodes and edges.
merge	git merge	Three-way merge two .abhi branches against a common base. Conflicts surface as CONTRADICTS edges.
fsck	git fsck	Validate a .abhi file without importing it.
show	git show	Inspect a .abhi file's summary stats without loading it.
grep	git grep	Execute a saved or ad hoc query against a .abhi file.
load_abhi_chunks	—	Load only selected or query-relevant chunks from a .abhi file.

Legacy tool names (export_graph_backup, import_abhi, diff_abhi, merge_abhi, validate_abhi, inspect_abhi, query_abhi) are still accepted and automatically mapped to their canonical equivalents.

Vault

Tool	Description
export_markdown_vault	Export the graph as an Obsidian-compatible Markdown vault.
import_markdown_vault	Import an edited Obsidian vault back into the graph non-destructively.

---

What To Ask The Agent

Ask the agent...	Tool called
"Remember that..."	observe_conversation
"What do you know about X?"	query_graph
"What changed recently?"	graph_diff
"Summarize context for a new session"	prime_context
"Show all stored topics"	get_topics
"Export my memory to a file"	commit
"Are there any duplicate nodes?"	dedup_candidates
"What's the quality of my graph edges?"	edge_quality_report

Edges are what make graph memory work. observe_conversation and decompose_and_store create edges automatically. If you only call store_node, you get isolated facts — not a connected graph.

For broad summarization tasks, prefer aggregate_graph over query_graph when you want a large scoped slice of memory instead of high-precision semantic ranking.

---

Graph Data Model

Node types: fact, entity, concept, preference, decision, question, note

Edge types: relates_to, contradicts, depends_on, part_of, updates, derived_from, similar_to

Temporal validity: Every node supports valid_from and valid_to fields. query_graph and aggregate_graph exclude expired nodes by default. Pass include_invalidated: true to include them, or as_of: "<ISO-8601 datetime>" to query the graph at a specific point in time. resolve_conflict with a winner node ID automatically sets the losing node's valid_to to now.

---

Cross-Client Handoffs & Migration

Same machine — automatic sharing

Point multiple clients at the same WAGGLE_DB_PATH (default ~/.waggle/memory.db) and they share one brain automatically.

Session handoffs

# Explicit checkpoint before switching sessions or apps
waggle-mcp checkpoint-context --project MCP --session-id thread-123 --output ./handoff.abhi

Resume order is:

• same machine / shared WAGGLE_DB_PATH: use the live SQLite memory first
• if that scoped DB recall is empty: import the session .abhi checkpoint
• different machine or explicit transfer: waggle-mcp pull ./handoff.abhi

Full migration

# Export
waggle-mcp export-graph-backup --output-path my_memory.json

# Import on new machine
waggle-mcp import-graph-backup --input-path my_memory.json

---

CLI Command Reference

Command	Description
waggle-mcp --help	Show all commands and options.
waggle-mcp features	Best first command — tool map, workflows, and setup hints.
waggle-mcp doctor	Run this if something isn't working — checks config, model cache, DB path.
waggle-mcp doctor --fix	Re-embed stale rows after a WAGGLE_MODEL change.
waggle-mcp setup --yes	Non-interactive one-line setup for all detected clients.
waggle-mcp init	Interactive setup wizard for one client.
waggle-mcp serve	Run the MCP server (usually started by your client).
waggle-mcp demo	Run the 60-second local demo with a pre-loaded example graph.
waggle-mcp edit-graph	Launch Graph Studio in the browser.
waggle-mcp uninstall-hooks	Remove the waggle-managed hooks block from Claude Code settings.
waggle-mcp export-context-bundle	Export a portable Markdown/JSON context pack.
waggle-mcp export-markdown-vault	Export the graph as an Obsidian-style vault.
waggle-mcp ingest-transcript-handoff	Ingest a rollover transcript, export a handoff bundle, and emit a session .abhi checkpoint.

WAGGLE_STARTUP_MODE

Value	Behaviour	Best for
normal (default)	Model loads in background; server responds immediately	Daily use
fast	ML never loads; semantic tools return unavailable	Schema inspection, tool listing
strict	Server blocks until model is fully loaded	Production deployments

---

Graph Studio

waggle-mcp edit-graph

A local browser-based graph editor for inspecting and editing memory directly. Features:

• Dual-layer graph/conversation views in the same UI
• Transcript provenance and retrieval-debug inspection for hybrid memory results
• Mouse-based node dragging and shift-drag edge creation
• Collapsible side panels, focus mode, and label toggling for large graphs
• Live graph stats: connected nodes, isolates, cluster count
• Export/import of the current graph, including .abhi preview, diff, and sharing workflows

.abhi files are JSON underneath — they support optional embedded vectors, optional AES-256-GCM encryption, deterministic content hashing, and a magic-bytes header (WGL\x01) for format identification. Legacy bare-ZIP files are read transparently.

waggle-mcp push encrypts .abhi exports by default. Export paths refuse to proceed if transcript text contains likely secrets unless you pass --force.

---

Model Support

Waggle uses a local sentence-transformers model selected by WAGGLE_MODEL.

• Default: all-MiniLM-L6-v2
• Any locally available sentence-transformers model name works.
• If the model is unavailable, Waggle falls back to deterministic SHA-256 embeddings.

WAGGLE_MODEL=all-mpnet-base-v2 waggle-mcp serve

For existing DBs, a model change is a migration event: run waggle-mcp doctor, then waggle-mcp doctor --fix if mixed embedding_model_id values are reported.

WAGGLE_DEDUP_THRESHOLD

Controls the cosine similarity threshold for automatic node deduplication at write time (default 0.88, minimum 0.85). Nodes above this threshold with matching type and scope are merged automatically. Use dedup_candidates to review near-duplicates below the auto-merge threshold.

---

Security & Privacy

• Data stays local by default (~/.waggle/memory.db). No telemetry, no cloud calls for local operation.
• Memory only leaves your machine if you configure a remote backend or explicitly export/push.
• Local SQLite is not encrypted at rest — use OS disk encryption if the stored history is sensitive.
• Before .abhi export, Waggle scans transcript text for likely secrets (API keys, JWTs, passwords). Export is refused if secrets are found unless you pass --force.
• waggle-mcp push defaults to AES-256-GCM encrypted export.

---

Known Limitations

• Edges are load-bearing. observe_conversation and decompose_and_store create them automatically. Raw store_node calls without follow-up edges produce disconnected nodes with no traversal value.
• Graph retrieval trades tokens for reasoning context. Factual lookups are often cheaper than chunked RAG; graph-expansion queries intentionally spend more tokens to carry update chains and contradictions.
• Hybrid rerank is not the default. The no-rerank hybrid path is stronger right now. The rerank path is available but intentionally not the launch default.
• Deduplication is similarity-based, not universal semantic equivalence. Broader production text may still require additional aliases or stricter domain guards.

---

Troubleshooting

Run waggle-mcp doctor first — it catches the most common issues automatically.

Symptom	Likely cause	Fix
waggle-mcp: command not found	pipx bin dir not on PATH	pipx ensurepath then restart terminal
store_node / query_graph hangs	Embedding model downloading (~420 MB)	Set WAGGLE_MODEL=deterministic for instant offline mode
UnicodeEncodeError: 'charmap' codec	Windows stdout not UTF-8	python -X utf8 or set PYTHONUTF8=1
Additional properties not allowed ('user_text', 'assistant_text')	Old pre-v0.2 field names	Use user_message + assistant_response
Waggle registered but agent doesn't see it (Antigravity)	Wrong config file	Agent reads ~/.gemini/antigravity/mcp_config.json, not the VS Code extension file
ModuleNotFoundError: No module named 'mcp'	Wrong Python environment	Use pipx install waggle-mcp and run via waggle-mcp ...
pipx install fails building sentence-transformers	Python/OS wheel mismatch	Use Python 3.11+; upgrade: python3 -m pip install -U pip setuptools wheel

---

Reference & Docs

• Environment variables, full tool surface, admin commands, Docker setup: waggle-mcp --help or docs/reference.md
• Production deployment: deploy/kubernetes/README.md
• Operations and troubleshooting: docs/runbooks/
• Automatic memory rules (copy-pasteable): docs/automatic-memory-rules.md
• Hook integration details: docs/hooks.md
• .abhi format spec: docs/abhi-format-v2.md

---

Contributing

This repository is maintained privately. Internal contributors can use the docs in this repo as the source of truth.