Welcome to Anchor
The Control Plane for Agent Identity - Governance, Audit, and Compliance
Control Plane for Agent Identity
Models change. Identity shouldn't.
Anchor is not just memory storage. It's the control plane that makes agent memory safe, auditable, compliant, and reversible. Unlike Mem0 or Zep, Anchor focuses on governance over storage. Identity governs memory: it determines what an agent may remember, how that memory evolves, and what must remain invariant across model swaps, backend changes, and time.
Get started in 3 steps (Identity-First Workflow):
- Create an Identity - Define your agent's persona, behavior, and memory policies
- Configure Policies - Set up PII handling, retention rules, and governance policies
- Create Memories - Start storing memories governed by your identity's policies
Why Identity-First? Your agent's identity determines what it can remember, how it behaves, and what policies apply. Start with identity to ensure proper governance from day one.
Memory Overview
Manage and explore your agent's memories, snapshots, and search capabilities
Memory is how your AI agent remembers context, preferences, and past interactions.
Memory features include:
- 🧠 Memories: Store individual facts, preferences, and context
- ⌕ Semantic Search: Find memories using natural language queries
- ◷ Snapshots: Capture point-in-time states of all memories
- ▣ QA Metrics: Monitor memory quality and effectiveness
Browse, create, edit, and delete individual memories. Organize by type, status, or search.
Search your memories using natural language queries. Find relevant context instantly.
Capture and restore complete memory states. Create snapshots for version control and rollback.
Track memory quality, confidence scores, and effectiveness metrics. Identify issues early.
Memory Management
Store, organize, and manage agent memories with enterprise-grade controls
Semantic Search
Search memories using natural language queries. Semantic search requires embeddings to be provided by your application—Anchor never generates embeddings to maintain LLM-agnostic architecture.
Snapshots
Create point-in-time backups and restore previous states
What are snapshots?
Snapshots save the current state of all memories for this subject at a specific point in time. You can later restore memories to any snapshot, which is useful for:
- Testing changes without losing data
- Recovering from incorrect memory updates
- Debugging memory-related issues
- Creating checkpoints before major changes
Audit Ledger
Complete audit trail of all memory operations for compliance and debugging
Operation Types:
- ADD: A new memory was created and stored
- UPDATE: An existing memory was modified
- DELETE: A memory was deleted (soft delete by default)
- NOOP: No Operation - A memory was attempted but blocked/quarantined by policy or eval gates (contradiction check, low confidence, PII detection, etc.)
- ROLLBACK: Memory state was restored to a previous snapshot
NOOP entries show that a memory write was attempted but rejected. Check the reason field to see why it was blocked.
Decision Logs
Track which memories influenced agent decisions for transparency and explainability
Decision logs are automatically created when you retrieve memories. They provide transparency into which memories influenced agent decisions. They show:
- Query: What was searched for
- Retrieved Memories: Which memories were found (with relevance scores if semantic search was used)
- Reasoning Trace: Step-by-step explanation of how the decision was made
- Used Memories: Which memories actually influenced the final response
Note on Semantic Search: Anchor supports semantic search when embeddings are provided by your application. Anchor never generates embeddings itself—this keeps the system LLM-agnostic. If embeddings are provided, decision logs will include relevance scores. If not, decision logs still track which memories were retrieved.
This provides full transparency into why your agent made specific decisions based on its memory.
Settings
Configure your workspace and subject for this session
Your session settings determine which workspace and subject you're working with. All operations (create, search, delete) will use these values.
- Workspace ID: Your organization or workspace identifier
- Subject: The user or entity whose memories you're managing (e.g., "user:123", "agent:bot-1")
Changes are saved automatically and persist across page refreshes.
Your current API key is the one used automatically when you log in with your email and password.
- One API key is marked as "Current" (shown with a blue badge)
- When you log in, the current key is used for your session
- You can change which key is current by clicking "Set as Current" and entering that key's value
- API keys are created only at signup - login uses your existing current key
What we don't track: Memory content, personal information, API keys, passwords
Memory Policies
Configure workspace-level default policies for memory storage, retention, and PII handling
Policies control what memories can be stored, how long they're retained, and how PII is handled. These settings apply to all memories created without a specific identity policy.
Identity
Manage your AI identity: personality, memory policies, and model configuration
Identity is your AI persona that persists across models, sessions, and deployments.
Each identity comprises:
- 🎭 Persona Rules: Behavioral guidelines that define how the identity responds and interacts
- 🌡️ Temperature: Controls creativity and randomness in responses
- 💾 Memory Policy: Governs what types of memories are stored (profile, episodic, semantic, procedural), retention periods, and PII handling
- 🔧 Model Policy: Specifies which AI model and provider to use (optional - defaults to workspace settings)
Key Insight: Models change. Your identity shouldn't. When you swap from GPT to Claude, your identity's persona rules, memories, and behavior stay consistent.
Model Configuration
Configure which AI model this identity uses
Model configuration tracks which AI model your identity uses.
This enables:
- Model Swap Tracking: See when you switch from GPT-4 to Claude and how it affects behavior
- Consistency: Your identity's personality and memories persist across model changes
- Audit Trail: Track which model was used for each interaction
Key Insight: Models change. Your identity shouldn't. When you swap models, your identity's personality and memories stay the same.
Memory QA Metrics
Quality metrics for agent memory: adherence, contradictions, and repetition
QA Metrics help you measure memory quality and agent behavior.
- Repeated Question Rate: How often the same questions are asked (proxy for memory effectiveness)
- Preference Adherence: How well the agent follows stored preferences (based on profile memory confidence)
- Contradiction Rate: How often new memories conflict with existing ones
Lower is better for repeated questions and contradictions. Higher is better for preference adherence.