agmem

agmem - Agentic Memory Version Control System

Git for AI Agent Memories

agmem is a version control system specifically designed for AI agent memory artifacts. It brings Git's proven concepts—commits, branches, merging, and distributed collaboration—to the world of agent episodic logs, semantic knowledge, and procedural workflows.

Why agmem?

AI agents accumulate vast amounts of memory: conversation histories, learned facts, user preferences, and problem-solving strategies. But unlike code, agent memory lacks:

• Version history - See what an agent learned and when
• Branching - Experiment safely with different memory strategies
• Merging - Combine knowledge from multiple agents
• Collaboration - Share and fork agent memories
• Rollback - Undo when agents learn the wrong things

agmem solves all of these problems with a familiar Git-like interface.

Features

🔧 Core Version Control

• ✅ Git-like workflow — init, add, commit, status, log, branch, checkout, merge, diff, show, reset, tag, stash, reflog, blame, tree, clean
• ✅ HEAD~n resolution — Walk parent chain: agmem log HEAD~5, agmem show HEAD~1
• ✅ Branch/tag names with / — Git-style refs: feature/test, releases/v1 (path-validated)
• ✅ Content-addressable storage — SHA-256 deduplication like Git with zlib compression
• ✅ Memory-type-aware merging — Episodic append, semantic consolidate, procedural prefer-new

🌐 Collaboration & Remotes

• ✅ Remote operations — clone, push, pull, remote with file:// URLs
• ✅ Multi-agent trust — Trust store (full/conditional/untrusted) per public key
• ✅ Federated collaboration — Coordinator API for distributed memory sharing
• ✅ Conflict resolution — agmem resolve with ours/theirs/both strategies

🔒 Security & Privacy

• ✅ Cryptographic verification — Merkle tree over blobs, Ed25519 signing
• ✅ Encryption at rest — AES-256-GCM with Argon2id key derivation
• ✅ Tamper-evident audit — Append-only hash-chained log
• ✅ Differential privacy — Epsilon/delta budget with fact-level noise
• ✅ Zero-knowledge proofs — Keyword containment, memory freshness
• ✅ PII scanning — Pre-commit hooks for sensitive data

🧠 Intelligence & Search

• ✅ Semantic search — Vector embeddings with GPU acceleration
• ✅ Knowledge graph — Wikilinks, tags, co-occurrence relationships
• ✅ Multi-provider LLM — OpenAI and Anthropic integration
• ✅ Temporal queries — Point-in-time and range-based retrieval

⚙️ Operations & Performance

• ✅ Pack files & GC — Object packing with 5-10x delta compression
• ✅ Health monitoring — Storage, redundancy, staleness, graph consistency
• ✅ IPFS/S3/GCS remotes — Distributed storage backends
• ✅ GPU acceleration — CUDA/MPS detection for embeddings

---

🚀 New in v0.3.0: Complete Feature Set

Phase 1: UX Parity

graph LR
    subgraph Daemon ["🔄 Observation Daemon"]
        D1["Extract MCP<br/>observations"] --> D2["Classify by<br/>memory type"]
        D2 --> D3["Auto-stage<br/>changes"]
        D3 --> D4["Generate commit<br/>message (LLM)"]
    end
    
    subgraph WebUI ["🌐 Web Viewer"]
        W1["Dashboard"] --> W2["Timeline"]
        W2 --> W3["Graph"]
        W3 --> W4["Agents"]
    end
    
    subgraph Search ["🔍 Progressive Search"]
        S1["Layer 1:<br/>Fast index"] --> S2["Layer 2:<br/>Timeline context"]
        S2 --> S3["Layer 3:<br/>Full details"]
    end
    
    subgraph Sessions ["📋 Session Manager"]
        SS1["Start session"] --> SS2["Track observations"]
        SS2 --> SS3["Classify topics"]
        SS3 --> SS4["Auto-commit"]
    end
    
    style Daemon fill:#e3f2fd,stroke:#1976d2,stroke-width:2px
    style WebUI fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
    style Search fill:#e8f5e9,stroke:#388e3c,stroke-width:2px
    style Sessions fill:#fff3e0,stroke:#f57c00,stroke-width:2px

Phase 2: Differentiation

graph TB
    subgraph Collab ["👥 Multi-Agent Collaboration"]
        C1["Agent Registry<br/>register, list, trust"]
        C2["Trust Manager<br/>grant, revoke, verify"]
        C3["Contribution Tracker<br/>commits, leaderboard"]
        C4["Conflict Detector<br/>concurrent edits"]
    end
    
    subgraph Compliance ["📋 Compliance Dashboard"]
        CP1["Privacy Manager<br/>budget tracking"]
        CP2["Encryption Verifier<br/>status checks"]
        CP3["Tamper Detector<br/>Merkle verification"]
        CP4["Audit Analyzer<br/>pattern detection"]
    end
    
    subgraph Archaeology ["🏛️ Memory Archaeology"]
        A1["History Explorer<br/>file evolution"]
        A2["Forgotten Finder<br/>unused memories"]
        A3["Pattern Analyzer<br/>recurring themes"]
        A4["Context Reconstructor<br/>historical state"]
    end
    
    subgraph Confidence ["📊 Confidence Scoring"]
        CS1["Decay Model<br/>exponential, linear, step"]
        CS2["Source Tracker<br/>reliability scores"]
        CS3["Confidence Calculator<br/>multi-factor scoring"]
        CS4["Expiration Alert<br/>low-confidence items"]
    end
    
    style Collab fill:#e1f5fe,stroke:#01579b,stroke-width:2px
    style Compliance fill:#fce4ec,stroke:#c2185b,stroke-width:2px
    style Archaeology fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
    style Confidence fill:#e8f5e9,stroke:#388e3c,stroke-width:2px

Phase 3: Advanced Features

graph TB
    subgraph TimeTravel ["⏰ Time-Travel Debugging"]
        T1["TimeExpressionParser<br/>'yesterday', '2 weeks ago'"]
        T2["TemporalNavigator<br/>commit lookup by time"]
        T3["TimelineVisualizer<br/>activity heatmaps"]
        T4["SnapshotExporter<br/>JSON, Markdown, Archive"]
    end
    
    subgraph PrivateSearch ["🔐 Privacy-Preserving Search"]
        P1["SearchTokenizer<br/>blind search tokens"]
        P2["AccessControl<br/>file-level permissions"]
        P3["DP Noise<br/>result randomization"]
        P4["Private Engine<br/>secure retrieval"]
    end
    
    subgraph SemanticGraph ["🕸️ Semantic Memory Graph"]
        G1["GraphBuilder<br/>auto-infer relationships"]
        G2["Clusterer<br/>type, tag, community"]
        G3["GraphSearch<br/>traversal queries"]
        G4["Visualizer<br/>D3 force-directed"]
    end
    
    subgraph Agents ["🤖 Memory Agents"]
        AG1["ConsolidationAgent<br/>merge fragmented"]
        AG2["CleanupAgent<br/>duplicates, stale"]
        AG3["AlertAgent<br/>notifications"]
        AG4["AgentManager<br/>orchestration"]
    end
    
    style TimeTravel fill:#fff3e0,stroke:#e65100,stroke-width:2px
    style PrivateSearch fill:#e8eaf6,stroke:#3f51b5,stroke-width:2px
    style SemanticGraph fill:#e0f2f1,stroke:#00695c,stroke-width:2px
    style Agents fill:#fbe9e7,stroke:#bf360c,stroke-width:2px

---

📊 Complete Architecture

flowchart TB
    subgraph Frontend ["🖥️ Frontend Layer"]
        direction LR
        CLI["CLI<br/>agmem *"]
        MCP["MCP Server<br/>30 tools"]
        WEB["Web UI<br/>React + WebSocket"]
        API["REST API<br/>22 endpoints"]
    end
    
    subgraph Core ["⚙️ Core Layer (48 Modules)"]
        direction TB
        
        subgraph VCS ["Version Control"]
            Repository
            Objects
            Refs
            Staging
            Merge
            Diff
        end
        
        subgraph Intelligence ["Intelligence"]
            VectorStore["Vector Store"]
            KnowledgeGraph["Knowledge Graph"]
            SemanticGraph["Semantic Graph"]
            SearchIndex["Search Index"]
            LLM["LLM Providers"]
        end
        
        subgraph Security ["Security"]
            CryptoVerify["Crypto Verify"]
            Encryption
            Trust
            Audit
            ZKProofs["ZK Proofs"]
        end
        
        subgraph Automation ["Automation"]
            Daemon
            Session
            Agents
            Gardener
            Distiller
        end
    end
    
    subgraph Storage ["💾 Storage Layer"]
        direction LR
        Local["Local<br/>.mem/objects/"]
        Pack["Pack Files<br/>Delta encoded"]
        IPFS["IPFS<br/>Distributed"]
        Cloud["S3/GCS<br/>Cloud"]
    end
    
    Frontend --> Core
    Core --> Storage
    
    style Frontend fill:#e3f2fd,stroke:#1976d2,stroke-width:3px
    style Core fill:#f5f5f5,stroke:#616161,stroke-width:2px
    style Storage fill:#fff3e0,stroke:#e65100,stroke-width:2px
    style VCS fill:#c8e6c9,stroke:#2e7d32
    style Intelligence fill:#bbdefb,stroke:#1565c0
    style Security fill:#ffcdd2,stroke:#c62828
    style Automation fill:#e1bee7,stroke:#7b1fa2

---

🔄 User Flows

Memory Commit Flow

sequenceDiagram
    participant Agent as 🤖 Agent
    participant Daemon as 🔄 Daemon
    participant Session as 📋 Session
    participant VCS as 💾 VCS
    participant Graph as 🕸️ Graph
    
    Agent->>Daemon: MCP tool observation
    Daemon->>Daemon: Extract & classify
    Daemon->>Session: Add to session
    Session->>Session: Topic classification
    
    alt Auto-commit enabled
        Session->>VCS: Stage changes
        VCS->>VCS: Generate commit (LLM)
        VCS->>Graph: Update relationships
    else Manual commit
        Agent->>VCS: agmem commit
        VCS->>Graph: Update relationships
    end
    
    Graph->>Graph: Rebuild connections

Multi-Agent Collaboration Flow

sequenceDiagram
    participant A1 as 🤖 Agent A
    participant Reg as 📋 Registry
    participant Trust as 🔐 Trust
    participant Remote as 🌐 Remote
    participant A2 as 🤖 Agent B
    
    A1->>Reg: agent_register
    A2->>Reg: agent_register
    
    A1->>Trust: trust_grant(Agent B)
    Trust->>Trust: Verify public key
    
    A1->>Remote: agmem push
    Remote->>Remote: Store commits
    
    A2->>Remote: agmem pull
    Remote->>Trust: Verify origin
    Trust->>A2: Merge if trusted
    
    A2->>A2: contribution_log

Time-Travel Debugging Flow

sequenceDiagram
    participant User as 👤 User
    participant Parser as 📅 Parser
    participant Nav as 🧭 Navigator
    participant VCS as 💾 VCS
    participant Export as 📦 Exporter
    
    User->>Parser: "2 weeks ago"
    Parser->>Parser: Parse expression
    Parser->>Nav: Get commit at time
    Nav->>VCS: Lookup by timestamp
    VCS->>Nav: Return commit hash
    
    alt View timeline
        Nav->>User: Activity heatmap
    else Export snapshot
        Nav->>Export: Export state
        Export->>User: JSON/Markdown/Archive
    end

---

Feature Coverage Overview

graph TB
    subgraph Core ["🔧 Core Features"]
        Core1["✅ Git-like commands<br/>init, add, commit, log, branch, merge"]
        Core2["✅ Content-addressable storage<br/>SHA-256, zlib, dedup"]
        Core3["✅ Memory-type merge<br/>Episodic, Semantic, Procedural"]
    end
    
    subgraph Collab ["👥 Collaboration"]
        Collab1["✅ Remote push/pull<br/>file:// URLs, conflict detection"]
        Collab2["✅ Multi-agent trust<br/>Trust store, key verification"]
        Collab3["✅ Federated sync<br/>Coordinator API"]
    end
    
    subgraph Safety ["🔒 Safety & Integrity"]
        Safety1["✅ Cryptographic<br/>Merkle, Ed25519"]
        Safety2["✅ Tamper detection<br/>Audit trail, hash-chain"]
        Safety3["✅ Encryption at rest<br/>AES-256-GCM"]
    end
    
    subgraph Privacy ["🕵️ Privacy"]
        Privacy1["✅ Differential privacy<br/>Epsilon/delta budget"]
        Privacy2["✅ Zero-knowledge proofs<br/>Keyword, Freshness"]
        Privacy3["✅ PII scanning<br/>Pre-commit hooks"]
    end
    
    subgraph Intelligence ["🧠 Intelligence"]
        Intel1["✅ Semantic search<br/>Vector + Text fallback"]
        Intel2["✅ Knowledge graph<br/>Wikilinks, Co-occurrence"]
        Intel3["✅ LLM integration<br/>OpenAI, Anthropic"]
        Intel4["✅ Temporal queries<br/>Point-in-time, Range"]
    end
    
    subgraph NewFeatures ["🆕 v0.3.0 Features"]
        New1["✅ Web UI + React Frontend<br/>Dashboard, Graph, Timeline"]
        New2["✅ Session Management<br/>Topic classification, auto-commit"]
        New3["✅ Memory Agents<br/>Consolidation, Cleanup, Alerts"]
        New4["✅ Time-Travel<br/>Natural language dates"]
    end
    
    subgraph Ops ["⚙️ Operations"]
        Ops1["✅ Pack/GC<br/>Object packing, cleanup"]
        Ops2["✅ Daemon mode<br/>Watch + auto-commit"]
        Ops3["✅ IPFS/S3/GCS<br/>Distributed storage"]
    end
    
    Core --> Collab
    Collab --> Safety
    Safety --> Privacy
    Privacy --> Intelligence
    Intelligence --> NewFeatures
    NewFeatures --> Ops
    
    style Core fill:#e3f2fd,stroke:#1976d2,stroke-width:2px
    style Collab fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
    style Safety fill:#ffebee,stroke:#c62828,stroke-width:2px
    style Privacy fill:#e1f5fe,stroke:#01579b,stroke-width:2px
    style Intelligence fill:#f1f8e9,stroke:#558b2f,stroke-width:2px
    style NewFeatures fill:#fff8e1,stroke:#f9a825,stroke-width:3px
    style Ops fill:#fff3e0,stroke:#e65100,stroke-width:2px

Quick Start

Installation

From PyPI (pypi.org/project/agmem):

pip install agmem

Or install from source:

git clone https://github.com/vivek-tiwari-vt/agmem.git
cd agmem
pip install -e .

Initialize a Repository

agmem init

This creates a .mem/ directory and a current/ working directory with subdirectories for different memory types:

.mem/                      # Repository internals
├── objects/               # Content-addressable storage
├── refs/                  # Branch and tag references
├── staging/               # Staging area
├── index.json             # Staging index
└── config.json            # Repository configuration

current/                   # Working directory
├── episodic/              # Session logs, conversation history
├── semantic/              # Facts, user preferences, knowledge
├── procedural/            # Workflows, prompt templates
├── checkpoints/           # Agent state snapshots
└── session-summaries/     # AI-generated summaries

Basic Workflow

# Add memory files
agmem add episodic/session-2026-01-31.md
agmem add semantic/user-preferences.md

# Or add all changes
agmem add .

# Check status
agmem status

# Commit changes
agmem commit -m "Learned user prefers dark mode and Python"

# View history
agmem log

# Create a branch for experimentation
agmem branch experiment
agmem checkout experiment

# Make changes and commit
agmem add .
agmem commit -m "Experimenting with new workflow"

# Switch back and merge
agmem checkout main
agmem merge experiment

# See who changed each line of a memory file (like git blame)
agmem blame current/semantic/user-preferences.md

# Stash work in progress, then restore
agmem stash push -m "WIP"
agmem stash list
agmem stash pop

# View HEAD history (reflog)
agmem reflog

Commands

All commands are listed below. Highlights: agmem blame <file> (who changed each line), agmem daemon (watch + optional auto-commit), agmem garden (episode archival & consolidation), agmem fsck (integrity check). See Inspection & maintenance for blame, daemon, and garden details.

Core (Git-like)

Remote & integrity

Optional (install extras)

Refs and optional extras

• Branch and tag names can include / (e.g. feature/test, releases/v1); the resolved path must stay under refs/heads or refs/tags (Git-style).
• HEAD~n is supported for walking the commit parent chain (e.g. agmem log HEAD~5, agmem show HEAD~1).
• Search: Semantic search when pip install agmem[vector]; otherwise agmem search falls back to plain text (grep-style) search over current/.
• Web: agmem serve and agmem graph --serve require pip install agmem[web].
• Daemon: agmem daemon requires pip install agmem[daemon].
• MCP: agmem mcp requires pip install agmem[mcp].

---

Inspection & maintenance

Blame — who changed each line

See which commit and author last modified each line of a memory file (like git blame):

agmem blame current/semantic/user-preferences.md
agmem blame episodic/session1.md

Output: commit hash (short), author, line number, and line content. Use this to trace where a fact or session line came from, or for audits.

Daemon — watch and auto-commit (optional)

Run a background process that watches current/ and can auto-commit when files change. Requires the daemon extra (watchdog).

pip install agmem[daemon]
agmem daemon run          # Watch current/, optional auto-commit after debounce
agmem daemon status       # Show if daemon is running
agmem daemon start        # Start in background (uses .mem/daemon.pid by default)
agmem daemon stop         # Stop the daemon

Use when you want memory changes to be committed automatically after a short idle period, or to keep a live view of the repo.

Garden — episode archival & consolidation (optional)

Archive old episodic files and consolidate semantic memory. All paths stay under current/ (path-validated).

agmem garden              # Run archival and consolidation (see config)
agmem garden --dry-run    # Show what would be archived/consolidated without writing

Configure in .mem/config.json (e.g. archive_dir, consolidation thresholds). Use to keep current/episodic/ manageable and to merge related semantic facts.

Integrity check

agmem fsck

Verifies objects, refs, and (if installed) the vector store. When commit metadata includes merkle_root and optionally signature, fsck also runs cryptographic verification. Run after cloning or if something looks wrong.

---

Security Architecture

agmem implements security in layers. Choose what you need:

graph TB
    subgraph L1 ["🔒 Tier 1: Cryptographic Integrity"]
        T1A["✅ Merkle tree<br/>over blobs"]
        T1B["✅ Ed25519<br/>signatures"]
        T1C["✅ Tamper<br/>detection"]
    end
    
    subgraph L2 ["🔐 Tier 2: Multi-Agent Trust"]
        T2A["✅ Trust store<br/>per public key"]
        T2B["✅ Multi-agent<br/>collaboration"]
        T2C["✅ Conflict<br/>resolution"]
    end
    
    subgraph L3 ["🕵️ Tier 3: Privacy & Anonymity"]
        T3A["✅ Differential<br/>Privacy budget"]
        T3B["✅ Zero-Knowledge<br/>Proofs"]
        T3C["✅ Encrypted<br/>at rest"]
    end
    
    subgraph L4 ["🌐 Tier 4: Distribution"]
        T4A["✅ S3/GCS<br/>remotes"]
        T4B["✅ IPFS<br/>support"]
        T4C["✅ Pack files<br/>& GC"]
    end
    
    subgraph L5 ["📊 Tier 5: Audit & Compliance"]
        T5A["✅ Append-only<br/>audit log"]
        T5B["✅ Hash-chained<br/>verification"]
        T5C["✅ Full history<br/>& blame"]
    end
    
    L1 --> L2 --> L3 --> L4 --> L5
    
    style L1 fill:#ffebee,stroke:#c62828
    style L2 fill:#ede7f6,stroke:#512da8
    style L3 fill:#e1f5fe,stroke:#01579b
    style L4 fill:#f3e5f5,stroke:#7b1fa2
    style L5 fill:#e8f5e9,stroke:#1b5e20

---

Security, trust & advanced features

The following 18 capabilities are implemented (or stubbed) per the agmem features implementation plan. They are grouped by tier.

Tier 1 — Security and trust

Tier 2 — Multi-agent collaboration

Tier 3 — Privacy

Tier 4 — Storage and distribution

Tier 5 — Intelligence and retrieval

Tier 6 — Operations and maintenance

New files and config (summary)

---

Memory Types

agmem recognizes three types of agent memory:

Episodic Memory

Time-stamped records of agent interactions:

current/episodic/
├── 2026-01-31-session1.md
├── 2026-01-31-session2.md
└── 2026-02-01-session1.md

Merge strategy: Append chronologically (no conflicts)

Semantic Memory

Factual knowledge and learned patterns:

current/semantic/
├── user-preferences.md
├── project-context.md
└── domain-knowledge.md

Merge strategy: Smart consolidation with conflict markers

Procedural Memory

Workflows and prompt templates:

current/procedural/
├── coding-workflow.md
└── debugging-process.md

Merge strategy: Prefer newer, flag for manual review

How It Works

Memory Flow

graph LR
    A["📂 Working Directory<br/>current/<br/>episodic/<br/>semantic/<br/>procedural/"] 
    B["📝 Staging Area<br/>index.json<br/>staged files"]
    C["💾 Object Storage<br/>.mem/objects/<br/>blobs → trees → commits"]
    D["📍 References<br/>.mem/refs/<br/>HEAD, branches, tags"]
    
    A -->|agmem add| B
    B -->|agmem commit| C
    C --> D
    
    style A fill:#e1f5ff,stroke:#01579b,stroke-width:2px
    style B fill:#fff3e0,stroke:#e65100,stroke-width:2px
    style C fill:#f3e5f5,stroke:#4a148c,stroke-width:2px
    style D fill:#e8f5e9,stroke:#1b5e20,stroke-width:2px

Merge Strategies by Memory Type

graph TB
    A["🌿 Branch A"] --> E{Memory Type?}
    B["🌿 Branch B"] --> E
    
    E -->|"<b>Episodic</b><br/>session logs"| F["⏰ Chronological<br/>Append"]
    E -->|"<b>Semantic</b><br/>learned facts"| G["🤝 Conflict<br/>Markers"]
    E -->|"<b>Procedural</b><br/>workflows"| H["✨ Prefer New<br/>with Flag"]
    
    F --> I["✅ No Conflicts<br/>Deterministic merge"]
    G --> J["⚠️ Manual Review<br/>Conflict markers"]
    H --> J
    
    style E fill:#ffebee,stroke:#b71c1c,stroke-width:2px
    style F fill:#c8e6c9,stroke:#1b5e20,stroke-width:2px
    style G fill:#ffe0b2,stroke:#e65100,stroke-width:2px
    style H fill:#b3e5fc,stroke:#01579b,stroke-width:2px
    style I fill:#81c784,stroke:#1b5e20,stroke-width:2px
    style J fill:#ffb74d,stroke:#e65100,stroke-width:2px

How Others Handle Memory vs agmem

graph LR
    subgraph Cursor ["Cursor<br/>Ephemeral"]
        C1["❌ No history"]
        C2["❌ No branching"]
        C3["❌ Session-based"]
    end
    
    subgraph Claude ["Claude Code<br/>File-based"]
        Cl1["⚠️ Manual Git"]
        Cl2["❌ Not memory-aware"]
        Cl3["⚠️ At launch"]
    end
    
    subgraph Mem0 ["Mem0<br/>Cloud/Vector"]
        M1["⚠️ Proprietary"]
        M2["❌ No branching"]
        M3["❌ Black box"]
    end
    
    subgraph agmem_box ["agmem<br/>Git for Memory"]
        A1["✅ Full history"]
        A2["✅ Branches & merge"]
        A3["✅ Local-first"]
        A4["✅ Type-aware"]
        A5["✅ Open source"]
    end
    
    style Cursor fill:#ffebee,stroke:#c62828,stroke-width:2px
    style Claude fill:#fff3e0,stroke:#f57c00,stroke-width:2px
    style Mem0 fill:#ede7f6,stroke:#512da8,stroke-width:2px
    style agmem_box fill:#c8e6c9,stroke:#1b5e20,stroke-width:3px

Example: Multi-Agent Collaboration

# Agent A creates base memory
agmem init --author-name "Agent-A"
echo "User prefers Python" > current/semantic/preferences.md
agmem add .
agmem commit -m "Initial preferences"

# Agent B clones and extends
# cd agent-b-memory
# agmem clone file:///path/to/agent-a
# echo "User also likes TypeScript" >> current/semantic/preferences.md
# agmem add .
# agmem commit -m "Added TypeScript preference"

# Agent A pulls and merges Agent B's changes
# agmem pull
# agmem merge agent-b/main

Storage (Git-Like)

agmem stores agent memory exactly like Git stores code:

Working directory: current/ with episodic, semantic, procedural, checkpoints, session-summaries. Same mental model as Git's working tree.

Security: Paths, refs, and object hashes are validated everywhere (no path traversal). Branch/tag names may include / but must resolve under refs/heads or refs/tags. Remote URLs (file://) and clone targets are validated. serve binds to 127.0.0.1 by default.

• Content-addressable objects — Blobs, trees, commits hashed with SHA-256
• Compression — zlib for efficient storage
• Deduplication — Same content = same hash, stored once
• Reflog — History of HEAD changes (commit, checkout)
• Stash — Temporary storage for work-in-progress

Configuration

Repository config (.mem/config.json)

Repository configuration is stored in .mem/config.json:

{
  "author": {
    "name": "Agent",
    "email": "agent@example.com"
  },
  "core": {
    "default_branch": "main",
    "compression": true,
    "gc_prune_days": 90
  },
  "memory": {
    "auto_summarize": true,
    "summarizer_model": "default",
    "max_episode_size": 1048576,
    "consolidation_threshold": 100
  },
  "differential_privacy": {
    "max_epsilon": 1.0,
    "delta": 1e-5
  },
  "federated": {
    "enabled": false,
    "coordinator_url": ""
  }
}

Optional sections: differential_privacy (for --private on distill/garden); federated (for agmem federated); signing (public key for commit verification); trust store under .mem/trust/ or config; encryption in .mem/encryption.json when enabled.

agmem config (cloud and PII)

Optional user/repo config for S3/GCS and PII hooks:

• User-level: ~/.config/agmem/config.yaml (or $XDG_CONFIG_HOME/agmem/config.yaml)
• Repo-level: <repo>/.agmemrc or <repo>/.mem/config.yaml (YAML)

Repo overrides user. Never put secrets in config files. Credentials are supplied via environment variables (or standard AWS/GCP mechanisms). Config may reference env var names only (e.g. access_key_var: AWS_ACCESS_KEY_ID). See docs/CONFIG.md for schema and examples.

• Cloud (S3/GCS): region, endpoint, env var names for credentials; GCS can use credentials_path (validated under repo/home) or credentials_json_var (env var containing JSON key).
• PII: pii.enabled (default true) to enable/disable pre-commit PII scanning; pii.allowlist (path globs) to skip scanning certain files.

Architecture

agmem follows Git's proven architecture:

graph TB
    subgraph Commands ["🎯 Commands - Porcelain (What you type)"]
        A1["init, add, commit, status, log, diff, show, tag, reset, tree"]
        A2["branch, checkout, merge, stash, clean, blame, reflog"]
        A3["clone, push, pull, remote, fsck, verify, audit, resolve"]
        A4["gc, prove, federated, search, graph, serve, daemon, garden"]
    end
    
    subgraph Core ["⚙️ Core Operations - Plumbing (What happens)"]
        B1["Staging Area<br/>index.json"]
        B2["Objects<br/>Blob, Tree, Commit, Tag"]
        B3["References<br/>HEAD, Branches, Tags"]
        B4["Memory-type Merge<br/>Episodic, Semantic, Procedural"]
    end
    
    subgraph Storage ["💾 Storage - Disk (Where data lives)"]
        C1["SHA-256 Hashing"]
        C2["zlib Compression"]
        C3["Deduplication"]
        C4["Validation<br/>Path, Ref, Hash"]
    end
    
    subgraph Features ["🔐 Advanced Features"]
        D1["Crypto Verification<br/>Merkle, Ed25519"]
        D2["Encryption at Rest<br/>AES-256-GCM"]
        D3["Privacy<br/>Differential Privacy"]
        D4["Trust & Audit<br/>Multi-agent, Hash-chain"]
    end
    
    Commands --> Core
    Core --> Storage
    Storage --> Features
    
    style Commands fill:#e3f2fd,stroke:#1976d2,stroke-width:2px
    style Core fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
    style Storage fill:#fff3e0,stroke:#f57c00,stroke-width:2px
    style Features fill:#e8f5e9,stroke:#388e3c,stroke-width:2px

Development

# Clone repository
git clone https://github.com/vivek-tiwari-vt/agmem.git
cd agmem

# Install development dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Format code
black memvcs/

# Type check
mypy memvcs/

Roadmap

• Core object storage (blob, tree, commit)
• Basic commands (init, add, commit, status, log, diff, show, reset, tag, stash, reflog, blame, tree, clean)
• HEAD~n resolution; branch/tag names with / (Git-style)
• Branching and checkout; merging with memory-type-aware strategies
• Remote operations (clone, push, pull, remote) — file:// URLs; pull merges into current branch; push conflict detection
• Search — semantic with agmem[vector], plain text fallback
• Knowledge graph (agmem graph) — wikilinks, tags, optional similarity; --no-similarity, --serve
• Integrity (agmem fsck); path/ref/hash validation; Merkle/signature verification
• Web UI (agmem serve); MCP server (agmem mcp); daemon (agmem daemon); garden (agmem garden)
• Cryptographic commit verification (Merkle tree, Ed25519 signing, verify on checkout/pull/fsck)
• Tamper-evident audit trail (agmem audit); multi-agent trust store; conflict resolution (agmem resolve)
• Encryption at rest (optional AES-256-GCM); differential privacy budget (--private on distill/garden)
• Pack files and garbage collection (agmem gc); ZK proofs and federated stubs (agmem prove, agmem federated)
• Multi-provider LLM (OpenAI, Anthropic); temporal range queries; daemon health checks; GPU detection; test suite and CI
• Phase 1: UX Parity — Observation daemon, web viewer UI, progressive disclosure search, session-aware auto-commit
• Phase 2: Differentiation — Multi-agent collaboration, compliance dashboard, memory archaeology, confidence scoring
• Phase 3: Advanced Features — Time-travel debugging, privacy-preserving search, semantic memory graph, memory agents
• React Frontend — Interactive force-directed graph visualization with search, zoom, pan, node dragging
• WebSocket Support — Real-time updates for file changes, commits, and agent activity
• Service Templates — systemd and launchd service files for daemon deployment
• IPFS remote (stub in place); full ZK circuits and federated coordinator

Integrations

  Cursor ────────╮
  Claude Code ───┼──►  MCP Server  ──►  current/  ──►  .mem/
  Windsurf ──────┤     (agmem mcp)      episodic/
  Continue ──────╯                      semantic/
                                         procedural/

  Python API ────►  Repository()  ──►  (same)
  LangChain
  LangGraph

Cursor

1. Install: pip install agmem[mcp]
2. Add MCP server to Cursor settings (Settings → MCP → Edit Config):

{
  "mcpServers": {
    "agmem": {
      "command": "agmem",
      "args": ["mcp"]
    }
  }
}

3. Use memory tools in chat: memory_read, memory_search, memory_add, memory_log, memory_diff
4. Optional: Add Cursor rules pointing to current/ for automatic context loading

Claude Code

Same MCP setup as Cursor. Claude Desktop uses MCP natively. File-based memory (CLAUDE.md, .claude/rules) can coexist with agmem—agmem adds version control on top.

Other IDEs

Any MCP-compatible client (Windsurf, Continue, etc.) works with the same config. Use the Python API for custom integrations.

Vector Search

pip install agmem[vector]
agmem search "user prefers Python"
agmem search "workflow" --rebuild   # Rebuild index after many changes

When agmem[vector] is not installed, agmem search uses plain text (grep-style) search over current/**/*.md and *.txt.

Knowledge graph

agmem graph                    # Summary: nodes, edges by type, isolated files
agmem graph --no-similarity     # Wikilinks + tags only (no vector store)
agmem graph --format json      # Full graph JSON
agmem graph --format d3 -o g.json   # D3 nodes/links for visualization
agmem graph --serve            # Serve graph UI (requires agmem[web])

Edges come from wikilinks ([[target]], [[target|label]]) and shared YAML frontmatter tags. Use --no-similarity when vector deps are not installed.

Web UI

pip install agmem[web]
agmem serve          # Browse history, view diffs in browser
agmem graph --serve  # Serve knowledge graph UI (same extra)

Use Cases

Multi-Agent Collaboration

• Agent A develops memory on main
• Agent B clones, branches, and extends: agmem clone file:///path/to/agent-a
• Agent A pulls and merges: agmem pull then agmem merge agent-b/main
• Memory-type-aware merge: episodic appends, semantic uses conflict markers, procedural prefers newer

Experiment Tracking

• Branch per experiment: agmem branch exp-v2-prompt
• Compare results: agmem diff main exp-v2-prompt
• Tag successful runs: agmem tag v1.0
• Roll back: agmem reset --hard HEAD~1

Compliance and Audit

• Full history: agmem log, agmem reflog
• Line-level attribution: agmem blame <file> — see which commit and author last changed each line (e.g. agmem blame current/semantic/user-preferences.md)
• Tamper-evident audit: agmem audit and agmem audit --verify for hash-chained operation log
• Cryptographic verification: agmem verify --crypto and agmem fsck for Merkle roots and Ed25519 signatures
• Visual audit: agmem serve for browser-based history and diff viewer
• Integrity: agmem fsck to verify objects, refs, and commit signatures

Ecosystem Plugin Patterns

LangChain / LangGraph

Use the agmem Python API or MCP tools as a memory backend:

from pathlib import Path
from memvcs.core.repository import Repository

repo = Repository(Path("."))
repo.stage_file("semantic/learned.md", content=b"Agent learned X")
repo.commit("Learned X")

Or connect via MCP and call memory_add from your agent.

MCP Resource URIs

• mem://current/semantic/user-preferences.md — Read memory file
• mem://current/episodic/session1.md — Session log
• mem://current/procedural/workflow.md — Procedural memory

Plugin Checklist

1. Detect agmem repo: .mem/ directory exists
2. Read from current/ for context
3. Write via agmem add + agmem commit or MCP memory_add
4. Optional: Use agmem search for semantic retrieval

Links

• PyPI: pypi.org/project/agmem — pip install agmem

GitHub repo (About): Use this description and topics in your repo’s About (gear → Description & topics):

• Description: Git for AI agent memories. Version control for episodic logs, semantic knowledge, and procedural workflows.
• Topics: ai agent memory version-control git vcs llm python

Contributing

Contributions are welcome! Please read our Contributing Guide for details.

License

MIT License - see LICENSE file for details.

Acknowledgments

agmem is inspired by Git's elegant design and the growing need for agent memory management in the AI ecosystem. Thanks to the Git team for proving that distributed version control can be both powerful and usable.

---

Made with ❤️ for AI agents and their developers