AMS: Audit Memory System

Durable control plane for multi-agent orchestration.

---

Audit-first agent memory and orchestration layer for MCP clients.

AMS gives AI assistants durable project scope, task control, proof-oriented execution trails, long-term memory, and orchestration primitives. Every action is auditable. Every decision is traceable. Every completion is provable.

Docs	lasteld.github.io/AMS
Tool Reference	docs/api/tool-reference.md
Architecture	docs/architecture/overview.md
Getting Started	docs/tutorials/getting-started.md
Obsidian Plugin	integrations/obsidian-plugin/README.md
Release Notes	RELEASE_NOTES.md

Control Plane v1.1.0

The AMS Control Plane provides a unified dashboard for system oversight and management:

Screen	Feature	Description
Dashboard	System Overview	Real-time vitals, task stats, roadmap progress, and health indicators
Tools	Tool Surface Manager	Browse, search, and inspect all 484 MCP tools with filtering by tier, domain, and capability
Audit	Cryptographic Audit Log	Tamper-evident action logs with Merkle tree verification and chain integrity checks
Agents	Agent Roster	Real-time view of active agents, their status, health, and current assignments
Workflows	Pipeline Overview	Visual GSD workflow execution with phase status, dependencies, and progress tracking
Memory	Memory Inspector	Search, browse, and manage memory frames with semantic search and cross-session bridging

Access via Obsidian plugin or programmatically through the AMS MCP tool surface.

Project Status

Metric	Value
Total Tasks	300
Completed	248 (82.7%)
In Progress	0
Backlog	52
Available Tools	484
Test Suite	121 passing
Roadmaps	8
Wave 1 (Runtime)	Complete
Wave 2 (Domains)	Complete
Wave 3-4 (Claude, Tests, Docs)	Backlog

Install

git clone https://github.com/LastEld/AMS.git
cd AMS
npm install
npm test
node src/server.js

MCP Client Configuration

{
  "mcpServers": {
    "ams": {
      "command": "node",
      "args": ["$AMS_ROOT/src/server.js"],
      "env": {
        "AMS_ROOT": "$AMS_ROOT",
        "AMS_WATCH_MODE": "none",
        "DOTENV_CONFIG_QUIET": "true"
      }
    }
  }
}

Replace $AMS_ROOT with your AMS installation path (e.g., /path/to/AMS on Unix or C:/path/to/AMS on Windows).

Optional: add ANTHROPIC_API_KEY to env to enable the Claude API tool surface (290 tools).

Verify with:

unified_init
unified_vitals
unified_help

Local Skills (Tier 1)

AMS includes 16 shared skills in .agents/skills/ (for Claude, Kimi, Codex) plus 2 Claude-specific skills in .claude/skills/.

PM and Orchestration

Skill	Purpose
ams-project-manager	PM and handoff controller
ams-tier1-chains	Meta-prompting tool chains for sub-agent spawning
ams-autonomous	Autonomous planning and apply flows

Task and Roadmap

Skill	Purpose
ams-task-management	Task CRUD with dependencies
ams-roadmaps-tasks	Roadmaps, tasks, and project sync
ams-roadmap-progress	Register roadmaps and update progress

GSD and Execution

Skill	Purpose
ams-gsd	GSD planning and orchestration
ams-gsd-execution	Workflow execution and agent spawning

Audit, Memory, and Proof

Skill	Purpose
ams-audit-memory	Audit, memory, and proof flows
ams-audit-proof	Audit sessions and verifiable proof artifacts
ams-memory-context	Memory packing and immutable context snapshots

Infrastructure and Integrations

Skill	Purpose
ams-mcp-server	MCP server operation and troubleshooting
ams-observability	Health, metrics, logs, and alerts
ams-obsidian-integration	Obsidian plugin operation
ams-growth-strategy	Growth strategy execution
repo-facing-polish	GitHub-facing cleanup and releases

See AGENTS.md for full skill documentation and coordination rules.

Tool Surface — 484 Tools

AMS exposes 484 MCP tools organized into two tiers. Every tool listed below is a real, callable MCP tool registered in src/controllers/index.js. No placeholders, no stubs.

Tier 1: Core AMS (194 tools)

The operational control plane. These tools manage projects, tasks, execution, proof, and memory.

Domain	Count	Tools	Purpose
Unified	12	unified_init, unified_vitals, unified_help, unified_stats, unified_set_project, unified_sync, unified_backup, unified_restore, unified_export, unified_import, unified_metrics, unified_architecture	System bootstrap, health, diagnostics
Tasks	10	task_create, task_update, task_delete, task_list, task_deps, task_next_actions, task_eisenhower, task_report, task_create_batch, task_link_roadmap, task_sync_to_gsd	Task CRUD, prioritization, dependencies
Roadmaps	16	roadmap_list, roadmap_get, roadmap_nodes, roadmap_node_get, roadmap_register, roadmap_validate, roadmap_validate_all, roadmap_repair, roadmap_repair_all, roadmap_search, roadmap_analyze, roadmap_export_to_gsd, roadmap_progress_get, roadmap_progress_update, roadmap_progress_update_batch, roadmap_next_topic, roadmap_prerequisites_check	Learning roadmaps, progress tracking
Context	7	context_create, context_get, context_latest, context_diff, context_history, context_verify, context_ensure	Immutable execution snapshots
Audit	8	audit_session_start, audit_session_end, audit_session_status, audit_session_stats, audit_session_set_context, audit_get_actions, audit_get_chain_info, audit_verify_chain, audit_analytics	Action logs, chain verification
Thought	6	thought_record, thought_plan, thought_tree, thought_verify, thought_decide, thought_for_action, thought_reasoning_trail	Plans, decisions, reasoning trails
Merkle	6	merkle_finalize, merkle_root, merkle_verify, merkle_proof, merkle_audit, merkle_attest	Cryptographic proof trees
Memory	13	memory_smart_store, memory_get, memory_forget, memory_stats, memory_verify, memory_embed, memory_semantic_search, memory_bundle, memory_pack, memory_map, memory_gc, memory_summarize, memory_cross_session, memory_prioritize, memory_retention, memory_tier_migrate, memory_context_compress	Persistent memory, semantic search
Analysis	10	analysis_search, analysis_similar, analysis_compare, analysis_recommend, analysis_keywords, analysis_cache, analysis_rag_index, analysis_rag_search, analysis_rag_stats	TF-IDF search, RAG, recommendations
GSD	22	gsd_plan, gsd_validate, gsd_state, gsd_dash, gsd_discuss, gsd_evolve, gsd_lease, gsd_skill, gsd_vitals, gsd_verify_comprehensive, gsd_report_generate, gsd_agent, gsd_agent_spawn, gsd_agent_status, gsd_agent_terminate, gsd_agent_pool_create, gsd_agent_pool_scale, gsd_parallel_map, gsd_phase_plan, gsd_phase_execute, gsd_phase_status, gsd_workflow_create, gsd_workflow_run, gsd_workflow_status, gsd_workflow_list, gsd_workflow_cancel	Orchestration, agents, workflows
Intelligent	18	intelligent_route, intelligent_learn, intelligent_predict, intelligent_classify, intelligent_match, intelligent_rank, intelligent_suggest, intelligent_optimize, intelligent_anomaly, intelligent_scale, intelligent_retry, intelligent_fallback, intelligent_detect_anomaly, intelligent_recommend, intelligent_analyze, intelligent_stats, intelligent_feedback, intelligent_router_analyze, intelligent_cache_stats, intelligent_cache_warm, intelligent_tune, intelligent_report	ML routing, anomaly detection
Observe	14	observe_health_check, observe_system_metrics, observe_latency_dashboard, observe_throughput_metrics, observe_error_rates, observe_log_analyze, observe_trace_flow, observe_alert_create, observe_alert_list, observe_alert_acknowledge, observe_audit_realtime, observe_claude_usage, observe_claude_cost, observe_token_tracking, observe_metrics_export, observe_report_generate	Metrics, tracing, alerting
Other Core	52	acl_* (4), alert_* (7), ams_autonomous_* (2), ams_session_* (3), analytics_* (14), notify (1), parallel_execute (1), performance_* (3), rag_* (3), watcher_* (3), webhook_* (3), workflow_run (1)	ACL, alerts, automation, analytics

Tier 2: Claude API Surface (290 tools)

Direct Anthropic API integration. Requires ANTHROPIC_API_KEY in env.

Domain	Count	Purpose
Core API	10	claude_chat, claude_complete, claude_embed, claude_stream, claude_models, claude_model_capabilities, claude_configure, claude_health_check, claude_health_check_detailed, claude_health_metrics
Vision	19	Image analysis, OCR, object detection, scene classification, face analysis, color extraction, chart reading
Batch Processing	14	Batch job creation, monitoring, results, webhooks for bulk API operations
Streaming	9	SSE and WebSocket streaming, multiplexing, flow control
Files & Documents	16	File upload/download, PDF extraction, document Q&A, entity extraction
Conversations	14	Multi-turn conversation management, search, analytics, archiving
Messages	9	Message storage, search, indexing, export
Workspaces	16	Workspace CRUD, members, sharing, comments, activity tracking
Organizations & Projects	10	Org management, project CRUD, templates
Budget & Billing	11	Budget tracking, alerts, forecasting, invoices
Usage & Analytics	20	Token usage, cost analysis, trends, forecasting, anomaly detection
Tool Chains	13	Tool composition, chaining, discovery, autocomplete
Citations	13	Source management, citation formatting, validation
Computer Use	15	Screen interaction, keyboard/mouse control, screenshot capture
Tokens	5	Token counting, analysis, optimization, validation
Context Management	8	Context compression, optimization, prioritization, statistics
Threads	4	Thread forking, merging, import/export
Extended Thinking	5	Thinking budget, content access, optimization, statistics
Prompt Cache	3	Cache creation, metrics, status
Resilience	8	Circuit breakers, bulkheads, fallbacks, rate limiting, deduplication
API Keys	4	Key creation, listing, revocation, usage tracking
Webhooks	10	Webhook registration, management, testing, logs
Templates	2	Template listing and application
Fine-tuning	1	Model fine-tuning
Cost Estimation	3	Cost estimation, projection, budget-vs-actual

Full tool reference with every parameter: docs/api/tool-reference.md

Architecture

MCP Client (Claude Code / Obsidian / Codex / Kimi)
    |
    v  stdio JSON-RPC
src/server.js ── MCP Server
    |
    v  middleware (auth, ACL, audit, rate-limit)
src/controllers/ ── Tool Router (484 tools)
    |
    +── src/domains/     (tasks, roadmaps, gsd, context, audit, memory, merkle, thought, architecture, retention, integrations)
    +── src/tools/       (claude-api, claude-advanced, claude-workspace, claude-vision)
    +── src/claude/      (Claude SDK wrappers: core, models, batch, streaming, resilience)
    |
    v
src/db/ ── SQLite (WAL mode, serialized writes)
    |
    v
data/ams.db ── Durable state (tasks, contexts, audit actions, thoughts, memory, merkle trees)

Client Matrix

Client	Config Location	Server Name
Claude Code (VSCode)	.vscode/mcp-settings.json or .mcp.json	ams-v5
Claude Desktop	%APPDATA%/Claude/claude_desktop_config.json	ams-v5
Codex	~/.codex/config.toml	ams
Kimi Code	~/.kimi/mcp.json	ams-v5
Obsidian	Plugin settings (spawns own server)	N/A

All clients point to the same node src/server.js entrypoint and same data/ams.db.

Obsidian Plugin (v1.0.0)

The Obsidian plugin provides a full-featured AMS Control Center with real-time visual task management, GSD workflow control, agent pool management, and live dashboards.

Key Features

Feature	Description
Graph View	Interactive visualization for tasks, projects, and sessions with expandable hierarchy
Bidirectional Sync	Real-time sync between Obsidian and AMS with conflict resolution
Task Tree Visualization	Hierarchical task view with status badges, progress bars, and context menus
Session Tracking	Audit-backed session notes with Merkle proof trails
Memory Frame Integration	Access and manage AMS memory frames directly from Obsidian
GSD Workflow Control	Start, pause, resume workflows and view phase status
Agent Pool Management	Monitor agent health, spawn/terminate agents, view utilization
Reporting Dashboard	Live SVG charts for task completion, agent metrics, and workflow timeline

Installation

Via BRAT (Recommended):

1. Install the BRAT plugin in Obsidian
2. Add LastEld/AMS with subpath integrations/obsidian-plugin
3. Enable the AMS plugin in Community Plugins

Manual Installation:

1. Download the latest release from GitHub Releases
2. Extract to <vault>/.obsidian/plugins/ams-obsidian/
3. Enable the plugin in Community Plugins

Configuration

Configure via Settings → AMS Obsidian:

Setting	Description	Example
amsRoot	Path to AMS installation	$AMS_ROOT or /path/to/AMS
defaultProject	Auto-connect to this project	ams-core
nodePath	Path to Node.js binary	node
autoStartServer	Connect on plugin load	true
syncInterval	Auto-sync interval (minutes)	5

See integrations/obsidian-plugin/README.md for full documentation.

Repository Layout

src/
  server.js              MCP server entrypoint (production)
  controllers/           Tool definitions and routing (40+ files)
  domains/               Business logic (tasks, gsd, context, audit, memory, merkle, etc.)
  tools/                 Claude API tool handlers
  claude/                Claude SDK wrappers (core, models, batch, streaming, resilience)
  db/                    SQLite schema, repositories, migrations
  middleware/            Auth, ACL, rate limiting, runtime locks
  config/                Runtime configuration (server, claude, security)
  watchers/              File system watchers
  metrics/               Observability and metrics collection
  security/              Security scanning and validation
  validation/            Zod schemas for input validation
  runtime/               Process coordination and daemon management
  websocket/             WebSocket server for streaming
integrations/
  obsidian-plugin/       Full Obsidian plugin v1.0.0 (MCP stdio, rate-limited sync, offline queue, graph view)
tests/
  unit/                  Unit tests
  integration/           Integration tests (121 total, all passing)
docs/                    GitHub Pages documentation
data/                    Runtime database and state
config/                  Configuration files

Proof Model

AMS implements a proof-grade execution trail:

1. audit_session_start — begin a tracked session
2. thought_plan — record the execution plan
3. context_ensure — snapshot the execution context
4. Execute work (task updates, code changes)
5. thought_record — record decisions and evidence
6. audit_verify_chain — verify the action chain integrity
7. merkle_finalize — generate cryptographic proof tree
8. merkle_root — retrieve the Merkle root hash

Every step is immutable once written. The Merkle tree provides tamper-evident proof that the recorded execution trail was not modified after the fact.

Operating Constraints

• SQLite is the durable store. All writes are serialized.
• MCP clients must share the same AMS_ROOT.
• Desktop clients: set AMS_WATCH_MODE=none (watchers are for dedicated runtimes).
• Proof-grade handoff requires audit + thought + merkle flows, not chat text.
• The server uses console.log = console.error to keep stdout clean for MCP JSON-RPC.
• Set DOTENV_CONFIG_QUIET=true in env to prevent dependency stdout pollution.

Models (Claude API)

Model ID	Name	Context	Max Tokens
claude-opus-4-6	Claude Opus 4.6	200K	16,384
claude-sonnet-4-6	Claude Sonnet 4.6	200K	16,384
claude-haiku-4-5-20251001	Claude Haiku 4.5	200K	8,192
claude-3-5-sonnet-20241022	Claude 3.5 Sonnet	200K	8,192
claude-3-5-haiku-20241022	Claude 3.5 Haiku	200K	8,192
claude-3-opus-20240229	Claude 3 Opus	200K	4,096
claude-3-sonnet-20240229	Claude 3 Sonnet	200K	4,096
claude-3-haiku-20240307	Claude 3 Haiku	200K	4,096

Default: claude-sonnet-4-6

License

Apache-2.0 WITH Commons-Clause