CarryMem — Your AI Finally Remembers Who You Are

Stop teaching AI who you are every single conversation.

Your portable AI memory — preferences, decisions, and corrections shared by MCP-compatible local clients.

Every time you open a new chat, you introduce yourself again. Your preferences, your decisions, your corrections — all forgotten. Switch from Cursor to Claude Code, from GPT to Claude, start from scratch every time.

You're not using AI. You're training it. Over and over.

CarryMem fixes this. It's a lightweight, zero-dependency memory system that stores who you are and makes that identity available to MCP-compatible local clients. Your AI remembers your preferences, your past decisions, and the corrections you've made — so you can focus on building, not repeating yourself.

English | 中文 | 日本語 | 한국어 | 繁體中文

Table of Contents

• The 30-Second Version

• What CarryMem Does

• Real User Scenarios

• Get Started

• 4 Reasons to Choose CarryMem

• How It Works

• Quick Start

  • Install

  • System Requirements

  • Dependencies

  • 5 Lines of Code

  • CLI (50+ commands)

• Core Features

  • Memory That Understands You

  • Preference Injection

  • Memory Lifecycle

  • Security & Portability

• Supporting Features

• Comparison

• PrefEval Benchmark

• Architecture

• Module Overview

• Advanced Usage

• Who Is This For?

• Documentation

• Project Status

• Contributing

• Citation

• License

Related MCP server: DataDam Personal Data MCP Server

🌟 The 30-Second Version

你每天见客户、开会、聊天，AI 问你一句你答一句，下次对话它又忘了你是谁。

CarryMem 让 AI 自动记住你的偏好和决策——不用每次重复说。配置一次，支持本地 stdio MCP 的客户端可以共享同一份记忆。

Technical users: see PrefEval benchmarks (83.0% ICLR 2025 Oral) and architecture docs below.

Topics: ai-memory mcp claude-code agent-memory cursor obsidian preference-injection sqlite llm-tools portable-memory

What CarryMem Does

5 scenarios you'll recognize:

"I don't want to tell AI my preferences every time" "I prefer PostgreSQL" "Use React not Vue" "No comments in code" — say it once, remembered forever.

"I switched AI tools and started from scratch" Taught AI in Cursor, now teaching it again in Claude Code. CarryMem makes your AI memory follow you.

"I want to take my data with me" Your AI memory is yours. One file to pack, restore on any machine, any tool.

"USB Carry — my memories in my pocket" 🔑 Pack your memories to an encrypted .carry file, copy to USB, unpack on a new machine. Your AI identity travels with you — preferences, decisions, corrections, and rules all intact. Every agent on the new machine instantly knows who you are.

"My team shares conventions across agents" Team lead packs company rules as a Skill bundle, every team member installs it. All agents enforce the same conventions — no more "I didn't know we use SSL."

🎯 Real User Scenarios

Scenario 1: The Multi-Tool Developer

Monday:  Tell Cursor "I prefer dark mode, PostgreSQL, React"
Tuesday: Open Claude Code — it already knows your stack
Friday:  Switch to TRAE — same preferences, zero repetition

How: carrymem setup-mcp --all --global configures the supported clients. Workbuddy uses the manual MCP configuration shown below. All configured clients can share one local memory database.

Scenario 2: The USB Carry — New Machine, Same Identity

1. On your laptop:  carrymem pack -o my_identity.carry --encrypt
2. Copy my_identity.carry to USB drive
3. At new workplace: Install CarryMem on new machine
4. carrymem unpack my_identity.carry
5. Every agent on the new machine knows your preferences, decisions, and rules

Encrypted + SHA-256 checksum — your identity is safe even if the USB is lost.

Scenario 3: The Team Lead

1. Create team conventions as rules: "Always use SSL", "Never deploy on Friday"
2. Pack as Skill: carrymem rules pack rules.json --name team-conventions
3. Share the .json file with team
4. Each member: carrymem rules install team-conventions.json --scope company
5. All agents now enforce company conventions automatically

Scenario 4: The Long-Term User

Month 1: "I prefer dark mode" → stored as user_preference
Month 3: "Switch to light mode" → auto-supersedes old preference
Month 6: carrymem whoami → shows "I prefer light mode" (dark mode archived)

Preferences evolve. CarryMem tracks the history.

Get Started (pick your path)

Using Cursor / Claude Code / TRAE?

pip install carrymem
carrymem init
carrymem setup-mcp --all --global
carrymem doctor

--all configures the clients supported by the setup CLI. Restart the client after configuration.

Using Workbuddy

Workbuddy supports the same local stdio MCP server, but CarryMem does not edit its configuration automatically. Add this entry to Workbuddy's MCP configuration file, usually ~/.workbuddy/mcp.json, and preserve any existing servers:

{
  "mcpServers": {
    "carrymem": {
      "command": "carrymem",
      "args": ["mcp"],
      "env": {
        "CARRYMEM_DATA_PATH": "/Users/your-name/.carrymem/memories.db"
      }
    }
  }
}

If carrymem is not on PATH, use the absolute Python executable that contains CarryMem:

{
  "command": "/path/to/python",
  "args": ["-m", "carrymem.integration.layer2_mcp"]
}

Restart Workbuddy after saving the configuration.

Verify it works (30 seconds)

In either client, ask:

Please call CarryMem's health_check and report whether the service is healthy.

Then ask:

Remember: I prefer PostgreSQL.

Start a new conversation and ask:

What database do I prefer? Please query CarryMem.

To verify cross-client synchronization, store the preference in one client and query it in the other.

Need to move your memory?

carrymem pack                    # Creates carrymem_identity_20260526.carry
# Copy to USB / cloud / new machine
carrymem unpack my_identity.carry  # All memories restored

# With encryption for sensitive data
carrymem pack -o my_memories.carry --encrypt   # Password-encrypted .carry file
carrymem unpack my_memories.carry              # Auto-detects encryption, prompts for password

Auto-backup & Recovery

carrymem backup                  # Manual backup (also auto-backup every 20 writes)
carrymem backup --list           # List all backups
carrymem backup --restore memories_backup_20260527_120000.db  # Restore from backup

📊 Academically Verified: CarryMem's preference injection accuracy (83.0%) was measured using the PrefEval protocol (ICLR 2025 Oral, Amazon Science), outperforming simple reminder (80.0%) and zero-shot (71.5%) baselines across 200 test items. See Citation below.

4 Reasons to Choose CarryMem

These are what make CarryMem different from every other memory solution:

1. Preference Injection Precision — 83.0% (Academically Verified)

• Measured by PrefEval (ICLR 2025 Oral, Amazon Science), 200-sample 3-condition comparison

• CarryMem 83.0% > simple reminder 80.0% > zero-shot 71.5%

• Proactive injection > full reminder — first system to prove this

• 24% fewer unhelpful responses than reminder (28 vs 38) — more precise, less noisy

2. Zero-LLM Classification — Rule-Based First

• The rule engine classifies eligible memories without LLM tokens; a historical internal snapshot reported 88% of sampled inputs on that path.

• Current reproducible classification behavior is covered by the rule-engine tests and the benchmark command below; the 88% figure has no versioned dataset artifact and is not a release gate.

• Historical P99 latency and competitor comparisons are reported separately unless accompanied by a reproducible artifact.

3. Lightweight & Portable — SQLite Only

• Zero external dependencies for core functionality

• Single .db file — carry your identity anywhere

• Works with MCP clients that support local stdio MCP servers

4. Industrial-Grade Engineering — 4907 Tests / mypy 0 / flake8 0

• 4907 tests passed, 10 skipped in the current full regression (optional vector/semantic dependencies account for the skips); E2E and TUI suites are included in the repository test run.

• mypy 0 errors across 160 source files — fully type-safe (CI blocking gate)

• flake8 0 errors — clean codebase, no lint violations (black + isort formatted)

• 24 sensitive-pattern redaction — auto-detects API keys, passwords, tokens before storage

• PatternAnalyzer God Class split (1547→171 LOC facade + 3 modules) — maintainable architecture

• 394 docstrings added — 50%→100% public API documentation coverage

• Maturity assessment: 80/100 (B) per 7-dimension DevSquad evaluation

How It Works

User Input → Auto-Classification (7 types, rule-based first) → Smart Storage (SQLite + FTS5)
    → Semantic Recall (cross-language) → Context Injection (token budget) → AI Tool

Quick Start

Install

pip install carrymem

Requires Python 3.12+. Check your version: python --version

From PyPI: https://pypi.org/project/carrymem/

For development: git clone https://github.com/lulin70/carrymem.git && cd carrymem && pip install -e ".[dev]"

System Requirements

• Python: ≥3.12 (64-bit)

• OS: macOS 10.15+, Ubuntu 20.04+, Windows 10+

• Disk: ~5MB for core, ~200MB with semantic search

• Memory: ~50MB base

Dependencies

| Multi-language | pycld2, langdetect | pip install carrymem[language] | | Semantic Search | sqlite-vec, sentence-transformers | pip install carrymem[semantic] | | Full (all features) | all above | pip install carrymem[full] | | Development | pytest, black, flake8... | pip install -e ".[dev]" |

Zero LLM dependency for core features — classification uses rule engine only.

Verify Installation

carrymem version

If command not found, add Python bin to PATH:

# macOS (add to ~/.zshrc)
export PATH="$HOME/Library/Python/3.12/bin:$PATH"

# Linux (add to ~/.bashrc)
export PATH="$HOME/.local/bin:$PATH"

# Or use Python module directly
python3 -m carrymem version

Then run carrymem doctor to check your setup.

5 Lines of Code

⚠️ Package vs Import Name: Install with pip install carrymem (lowercase), but import as from carrymem import CarryMem (CamelCase class). The package name (carrymem) and class name (CarryMem) differ in casing.

from carrymem import CarryMem

cm = CarryMem()
cm.classify_and_remember("I prefer dark mode")        # Auto-classified as preference
cm.classify_and_remember("Use PostgreSQL not MySQL")   # Auto-classified as correction
cm.classify_and_remember("I prefer light mode now", session_id="sess_002")  # Session-aware
memories = cm.recall_memories("database")              # Semantic recall
memories = cm.recall_memories("mode", filters={"session_id": "sess_002"})  # Filter by session
agg = cm.recall_aggregated()                           # Aggregate by type
timeline = cm.recall_timeline("database")              # Knowledge evolution
print(cm.build_system_prompt())                        # Inject into any AI
cm.close()

CLI (50+ commands)

carrymem init                           # Initialize
carrymem add "I prefer dark mode"       # Store a memory
carrymem add "test note" --force        # Force store (bypass classification)
carrymem list                           # List memories
carrymem search "theme"                 # Search memories
carrymem show <key>                     # View memory details, source, and version
carrymem edit <key> "new content"        # Edit a memory and create a version
carrymem forget <key>                    # Delete a memory
carrymem whoami                         # Who your AI thinks you are
carrymem profile export --output identity.json   # Export your AI identity
carrymem stats                          # Memory statistics
carrymem check                          # Quality & conflict check
carrymem clean --expired --dry-run      # Preview cleanup
carrymem doctor                         # Diagnose installation
carrymem setup-mcp --tool cursor        # One-line MCP config
carrymem tui                            # Terminal UI
carrymem export backup.json             # Export all memories
carrymem import backup.json             # Import memories
carrymem pack -o my_memories.carry      # Pack into portable .carry file
carrymem pack -o my_memories.carry --encrypt  # Encrypted .carry file
carrymem unpack my_memories.carry       # Unpack .carry file
carrymem backup                         # Manual backup
carrymem backup --list                  # List backups
carrymem backup --restore <file>        # Restore from backup
carrymem version                        # Show version
# Rule Engine commands
carrymem rules add "use SSL" --trigger "database" --type avoid  # Add a rule
carrymem rules list --status active                              # List active rules
carrymem rules suggest                                           # Suggest rules
# Also available: carrymem add-rule, carrymem list-rules (legacy aliases)
carrymem rules pack rules.json --name team-conventions   # Pack rules as Skill
carrymem rules install team-conventions.json --scope company  # Install Skill
carrymem rules verify team-conventions.json              # Verify Skill integrity

Core Features (powering the 3 advantages)

Memory That Understands You

Auto-Classification (7 Memory Types)

CarryMem automatically identifies what kind of information you're sharing:

Semantic Recall (Cross-Language)

cm.classify_and_remember("我偏好使用PostgreSQL")

# All of these find it:
cm.recall_memories("PostgreSQL")     # Exact match
cm.recall_memories("数据库")          # Synonym expansion
cm.recall_memories("Postgres")       # Spell correction
cm.recall_memories("データベース")    # Cross-language (Japanese)

Identity Layer (whoami)

identity = cm.whoami()
print(identity["preferences"])   # ["I prefer dark mode", ...]
print(identity["decisions"])     # ["Let's use React", ...]
print(identity["corrections"])   # ["The port should be 5432", ...]

$ carrymem whoami

  Who You Are (according to your AI)
  ==================================================

  Your Preferences:
    ⭐ I prefer dark mode for all editors
    ⭐ I use PostgreSQL for databases
    ⭐ I always use Python for data analysis

  Your Decisions:
    🎯 Let's use React for the frontend

  Your Corrections:
    🔧 The port should be 5432, not 3306

  Memory Profile:
    Total: 19 | Dominant: user_preference | Avg Confidence: 73%

Preference Injection (advantage #1)

Version Chain — Preferences Evolve, Old Versions Auto-Archived

cm.update_memory(key, "Updated content")     # Creates version 2
history = cm.get_memory_history(key)          # [v1, v2]
cm.rollback_memory(key, version=1)            # Restore v1

Scope-Aware Injection — Only Inject Relevant Preferences Per Context

Preferences are injected based on context scope, so your database preferences don't clutter frontend discussions.

Token Budget — 60% Budget for Preferences, Never Truncated

CarryMem allocates 60% of the token budget to preferences, ensuring they're never cut off. This is the key to achieving 83.0% on PrefEval — structured preference injection beats simple reminders.

Memory Lifecycle (advantage #2)

Importance Scoring — Confidence × Type × Recency × Access

Every memory has an importance score that evolves over time:

importance = confidence × type_weight × recency_factor × access_factor

• 30-day half-life decay — old memories fade unless accessed

• Access reinforcement — frequently recalled memories stay fresh

• Type weighting — corrections (1.3x) > decisions (1.2x) > preferences (1.1x)

Consolidation (P0/P1/P2) — Dedup + Decay + Pattern → Rules + Semantic Merge

Automatic memory lifecycle management with three phases:

# Preview what consolidation would do
report = cm.consolidate(dry_run=True)
print(f"Duplicates: {report['stats']['duplicates_found']}")
print(f"Decayed: {len(report['to_decay'])}")

# Run consolidation (P0: dedup + decay, P1: pattern promotion, P2: semantic merge)
report = cm.consolidate(dry_run=False, run_p1=True, run_p2=True)

Preferences are always preserved — never decayed or deduplicated.

Scheduled Consolidation — Automatic Background Maintenance

Run consolidation automatically on a recurring interval:

# Schedule consolidation every hour (runs in background thread)
cm.schedule_consolidation(interval_hours=1.0)

# Stop the scheduled consolidation
cm.stop_consolidation()

CLI:

carrymem consolidate --schedule 1h   # Run consolidation every hour
carrymem consolidate --stop          # Stop scheduled consolidation

Security & Portability (advantage #3)

Auto-Redaction — 24 Sensitive Patterns

Automatically detects and redacts API keys, passwords, tokens, and 21 other sensitive patterns before storage.

Encryption — AES-128 at Rest

cm = CarryMem(encryption_key="my-secret-key")
# All content encrypted at rest, decrypted on read

Backup/Restore — Auto-Backup + Manual Control

Auto-backup triggers every 20 write operations (VACUUM INTO), retaining up to 5 backup files. Manual control via CLI:

carrymem backup                  # Create manual backup
carrymem backup --list           # List all backups
carrymem backup --restore <file> # Restore from a specific backup

Pack/Unpack — USB Carry with Encryption

# Pack memories into a portable .carry file
carrymem pack -o my_memories.carry

# With password encryption for sensitive data
carrymem pack -o my_memories.carry --encrypt

# Unpack on any machine (auto-detects encryption)
carrymem unpack my_memories.carry

SHA-256 checksum ensures file integrity. v1.0 .carry format is backward compatible with a warning.

Export/Import — Identity Follows You Across Devices

# Export your AI identity
cm.export_profile(output_path="my_identity.json")

# On another device or AI tool
cm.import_memories(input_path="backup.json")

Supporting Features

Error Code System (v0.4.0 New)

Structured error handling with bilingual messages:

from carrymem.errors import CarryMemError

# Error code ranges:
# CM-001~099: Configuration & Initialization
# CM-100~199: Storage Adapter
# CM-200~299: Memory Operations
# CM-300~399: Classification & Rule Engine
# CM-400~499: Security & Encryption
# CM-500~599: Import / Export
# CM-600~699: CLI / TUI / MCP Entry Points

try:
    cm.classify_and_remember("test")
except CarryMemError as e:
    print(e.code)     # "CM-001"
    print(e.message)  # "存储适配器未配置。" (Chinese)
    print(e.hint)     # "💡 Use CarryMem(storage='sqlite')..."

Features:

• 37 error codes with Chinese + English messages

• from_cause() factory maps low-level exceptions → friendly codes

• Actionable hints for every error

• 7 concrete error subclasses for programmatic handling

Monitoring Framework

Metrics are recorded automatically by the real operation paths — no manual increment() calls are needed in application code:

from carrymem import CarryMem
from carrymem.monitoring import HealthChecker, get_metrics_collector

cm = CarryMem(storage="sqlite")
cm.classify_and_remember("I prefer dark mode")   # counts + times the operation
cm.recall_memories("dark mode")                  # counts + times the operation

# The process-wide collector both ends share:
metrics = get_metrics_collector()
print(metrics.to_prometheus())                   # Prometheus text format

# SLO evaluation runs over the same data
health = HealthChecker(metrics_collector=metrics)
status = health.check()  # {"status": "ok"|"degraded", "checks": {...}, "slo": [...]}

Exposed series (authoritative producer table: docs/design/V0.11.0_OBSERVABILITY_INSTRUMENTATION.md):

In HTTP mode (carrymem serve) the same data is scraped from GET /metrics, and GET /healthz returns the SLO section (503 when a target is violated).

HTTP timeout contract: the request timeout ends client waiting only. A synchronous worker that already started may continue and may commit side effects. The JSON-RPC timeout response exposes error.data.execution=background_may_continue and error.data.side_effects=possible. The server drains started workers before closing CarryMem and SQLite. Clients must treat timed-out writes as potentially committed and use idempotent retry behavior.

SLO Targets:

• classify_and_remember P99 < 200ms

• recall P99 < 500ms

• startup < 2s — from import carrymem to the first CarryMem() construction, so it covers module import plus the first construction; one sample per process. sentence_transformers/torch are imported lazily (only when vector search is actually enabled), so they are not part of startup.

Honest boundaries:

• Metrics are per process and reset on restart — CarryMem is a local single-instance tool, so no cross-process aggregation is attempted.

• "status": "no_data" in the SLO section means the operation has not run in this process yet. If it runs and still shows no_data, that is an instrumentation bug, not health.

• sentence_transformers is probed lazily: importing carrymem does not import it or torch. Enabling vector search (enable_vector_search=True with the vector extra installed) still loads the embedding model during SQLiteAdapter construction — that is the explicit cost of vector search, not an import side effect.

Permission System (v0.4.0 New)

Lightweight access control MVP:

from carrymem.security.permissions import Permission, AccessPolicy

policy = AccessPolicy(owner_id="user-123")

# Check permissions
policy.check("user-123", Permission.READ)   # True
policy.check("other-user", Permission.WRITE)  # False

# Require permission (raises SecurityError on denial)
policy.require("user-123", Permission.DELETE, resource="memory")

i18n Internationalization (v0.4.0 New)

Multi-language support without gettext dependency:

from carrymem.i18n import I18nManager, set_locale, _

# Switch language
set_locale("zh-CN")

# Translate with variable interpolation
print(_("memory.stored", count=3))
# → "已记住 3 条记忆"

# Available locales: en, zh-CN

MCP Integration (One-Line Setup)

# Configure for Cursor
carrymem setup-mcp --tool cursor

# Configure for Claude Code
carrymem setup-mcp --tool claude-code

# Configure for all
carrymem setup-mcp --tool all

35 MCP tools available, including memory update, version history, pause and resume controls. Use tools/list or carrymem doctor to verify the installed server.

Client Compatibility:

The third-client real UI acceptance is intentionally skipped in the current local rollout. Do not treat protocol-level compatibility as client UI acceptance.

🔒 Your memories stay on your machine. CarryMem stores all data locally in ~/.carrymem/ (SQLite). Each user gets an independent database — just like Git, everyone uses the same tool but keeps their own repos. No cloud sync, no shared state, no cross-user conflicts.

Rule Engine with Scopes

Behavioral rules with three scope levels for team/organization alignment:

from carrymem.rules import RuleEngine

engine = RuleEngine()

# Company-mandated rules (highest priority, cannot be overridden)
engine.add_rule("database", "Always use SSL", scope="company", override=True)

# Personal preferences (lowest priority)
engine.add_rule("database", "Prefer PostgreSQL", scope="personal")

# Scope-aware matching
results = engine.match("database design", scopes=["company"])

Skill Format — Portable Rule Bundles

Share rule sets across teams with cryptographic integrity:

# Pack rules into a portable Skill bundle
bundle = engine.skill_pack(
    name="team-conventions",
    version="1.0.0",
    scope="company",
    author="team-lead",
)

# Verify integrity before installing
result = engine.skill_verify(bundle)
assert result["valid"] is True

# Install on another machine
engine.skill_install(bundle, scope_override="company", mode="skip")

Merge Protocol — Conflict Resolution

Three strategies for merging rules from different sources:

Quality Management

carrymem check                    # Check all
carrymem check --conflicts        # Detect contradictions
carrymem check --quality          # Find low-quality memories
carrymem check --expired          # Find expired memories
carrymem clean --expired --dry-run # Preview cleanup

Terminal UI

pip install textual
carrymem tui

Interactive terminal interface with sidebar filters, search, add, delete (d), and edit (e).

VS Code Extension

Rule management directly in your editor:

• Rule sidebar with scope badges

• Add/edit/delete rules via webview

• Effectiveness report panel

• Skill pack/install from file dialogs

Comparison

By Scenario

Feature Matrix

Note: Comparison based on publicly available information. Products evolve rapidly — please verify latest features. See docs/design/METHODOLOGY.md for the full design space positioning diagram and orthogonal classification table.

Key Difference: Other products store what you read. CarryMem stores who you are.

Design Space Positioning

CarryMem occupies a unique position at the intersection of three axes:

Portable ←————————————————————————————————→ Team-shareable
   |          CarryMem ⭐                  Memobase
   |                                    (team, retrieval, external)
   |     Mem0
   |   (solo, retrieval, vector)
   |————————————————————————————————————————————→ Executable
   |                                    User as Code
   |                                    (solo, executable, Python)

CarryMem is the only framework at the intersection of team sharing, executable rules, and zero dependencies.

🏆 PrefEval — Preference Adherence Benchmark

Protocol: PrefEval (ICLR 2025 Oral, Amazon Science) Sample: 200 items, 10 inter-turns, Claude Sonnet 4

Why this matters: Reminder injects "remember user preference" in every turn. CarryMem injects structured preferences in system prompt — more precise, more persistent, 24% fewer unhelpful responses.

Architecture

Three-Layer: Mixin + Facade + Protocol (v0.4.0)

┌─────────────────────────────────────────────────────────┐
│                    Facade Layer                          │
│  CarryMem (unified entry point, health_check, version)  │
├─────────────────────────────────────────────────────────┤
│                   Mixin Layer (8 modules)                │
│  Lifecycle │ MemoryCRUD │ Classification │ Recall       │
│  Backup    │ ProfileExport │ Maintenance │ PromptDelegate│
├─────────────────────────────────────────────────────────┤
│                  Protocol Layer (10 Protocols)            │
│  LifecycleOps │ BackupOps │ RecallOps │ ... │ CarryMemOps│
└─────────────────────────────────────────────────────────┘

Data Flow:

User Input
    ↓
Auto-Classification (7 types, 4 tiers)
    ↓
Importance Scoring (confidence × type × recency × access)
    ↓
Smart Storage (SQLite + FTS5, WAL mode, thread-local pool, dedup, TTL, encryption)
    ↓
Memory Consolidation (P0: dedup+decay → P1: pattern→rules → P2: semantic merge)
    ↓
Semantic Recall (FTS5 + synonyms + spell fix + cross-language)
    ↓
Context Injection (token budget, relevance ranking)
    ↓
AI Tool (Cursor / Claude Code / any MCP client)

Three-Tier Classification:

Rule Engine (60%+) → Pattern Analysis (30%) → Semantic (10%)
     ↓                      ↓                      ↓
 Zero cost            Near-zero cost          Token cost

PatternAnalyzer Composition (v0.4.0 refactor — 1547→171 LOC):

PatternAnalyzer (facade, 171 LOC)
    ├── NoiseDetector          — noise filtering (B1-B5, C5 rules)
    ├── FeedbackDetector       — execution feedback detection
    └── MemoryPatternDetectors — 8 memory type detectors
        (preference/correction/fact/relationship/task/decision/sentiment/location)

Backward compatible: from carrymem.layers.pattern_analyzer import PatternAnalyzer unchanged.

Module Overview

Advanced Usage

Obsidian Knowledge Base

from carrymem import CarryMem, ObsidianAdapter

cm = CarryMem(knowledge_adapter=ObsidianAdapter("/path/to/vault"))
cm.index_knowledge()
results = cm.recall_from_knowledge("Python design patterns")

Async API

from carrymem import AsyncCarryMem

async with AsyncCarryMem() as cm:
    await cm.classify_and_remember("I prefer dark mode")
    memories = await cm.recall_memories("theme")

JSON Adapter (No SQLite)

from carrymem import CarryMem, JSONAdapter

cm = CarryMem(adapter=JSONAdapter(path="/path/to/memories.json"))

Memory Versioning

cm.update_memory(key, "Updated content")     # Creates version 2
history = cm.get_memory_history(key)          # [v1, v2]
cm.rollback_memory(key, version=1)            # Restore v1

Export Identity for Other AIs

# Export your AI identity
cm.export_profile(output_path="my_identity.json")

# On another device or AI tool
cm.import_memories(input_path="backup.json")

Who Is This For?

Tired of repeating yourself? You use Cursor, Claude Code, ChatGPT daily. You've told AI your stack, your style, your decisions a hundred times. And it still asks "what framework do you prefer?" CarryMem makes your AI remember — so you don't have to keep reminding it.

Maintaining CLAUDE.md by hand? You already know AI needs memory. You have prompt files everywhere. They conflict, they go stale, and they don't follow you between tools. CarryMem auto-classifies your preferences, decisions, and corrections — and keeps them fresh automatically.

Building AI agents? Your agents forget users between sessions. You need a memory layer that's lightweight, local, and works with any LLM. CarryMem gives you 5-line integration, 7 memory types, and a rule engine — with zero dependencies beyond SQLite.

Documentation

• Quick Start Guide

• Installation Guide

• User Guide

• Architecture

• API Reference

• API Stability Policy

• Roadmap

• Contributing

Project Status

Current Version: v0.11.2 Tests: 4907 passed, 10 skipped, 77 deselected in the current full regression; 265 E2E tests collected, with the latest realistic non-slow user-path run at 245 passed, 20 deselected, 2 warnings in 82.24s (the previous 77.08s run is retained as a historical baseline) Coverage: 83.19% mypy: 0 errors (160 source files, CI blocking gate) flake8: 0 errors (black + isort formatted) radon: 0 D/E/F functions (CI blocking gate) Blocking gates: eight local gates pass, including version-consistency and swallowed-assert Maturity: 80/100 (B) per 7-dimension DevSquad evaluation

Changelog:

• v0.11.2: Completes the recall metric correction — the coreference and correction-history reads inside core/_classification.py no longer count as user recalls, and the guard that missed them now covers four message shapes plus a control group. See CHANGELOG.md.

• v0.11.1: Fixes the release gate that blocked v0.11.0 (a steady-state SLO assertion ran on the first call of a process), corrects the first recall leak, and adds the full regression as a blocking local CI gate.

• v0.11.0: Breaking async API cleanup — AsyncCarryMem is the executor-backed async facade only; standalone native async SQLite I/O remains available through AsyncSQLiteAdapter. The core encryption floor is cryptography>=50.0.0. See CHANGELOG.md and ADR-009.

• v0.10.0: Repeat-correction upgrade — detect_repeat_correction() + semantic dedup (Jaccard + entity) + security-keyword bypass (forge 借鉴点 2). See docs/design/V0.10.0_REPEAT_CORRECTION.md.

• v0.9.9: Methodology: orthogonal classification table (docs/design/METHODOLOGY.md) + design space positioning (README Comparison + COMPETITIVE_ANALYSIS). See docs/design/METHODOLOGY.md.

• v0.9.8: Knowledge graph deletion completeness — forget() now cascades to memory_entities + memory_relations (TD-066, Oracle Agent Memory report启发). 4 new tests.

• v0.9.7: Tech debt cleanup — TD-003b/009/011b/002 follow-ups (deleted cli.py facade, added TUI fallback tests, downgraded SQLITE_SCHEMA to P3 observation).

• v0.9.6: Nightly slow test threshold fix (store_messages batch API) + release.yml cp consistency.

• v0.9.4: TD-063/TD-064/TD-065: test skip cleanup + flake8 bugbear fix + ruff config lock.

• v0.9.3rc1: TD-015 Password-based publish restored (OIDC abandoned).

• v0.9.2: P3 Tech Debt Cleanup — TD-031/049/050/051/053/054.

• v0.9.1: TD-055 — mypy src/ baseline errors cleared 9 → 0.

• v0.9.0: UI/UX Overhaul — Morandi Aesthetic + Accessibility + Onboarding.

• v0.8.0: Graphify — 3 new MCP graph tools (query_graph, shortest_path, get_memory_impact), edge confidence labels (EXTRACTED/INFERRED/AMBIGUOUS) on memory_relations, schema migration v100.

• v0.7.3: Security hardening — removed HMAC-CTR stream cipher fallback (cryptography is now a hard dependency), Fernet-only encryption, migration script for pre-v0.7.3 databases. Input validator defense-in-depth, batched LIKE queries, WAL throttle for recall access updates.

• v0.7.2: Native Async I/O — async_sqlite adapter (aiosqlite), async recall/store APIs, Memify enhancement. P0 security fixes (fail-closed access control, MCP dispatcher injection), CI/CD hardening (bandit blocking, pip-audit, Docker non-root, pre-release test gate).

• v0.7.1: Multi-Mode Retrieval — vector + FTS + semantic fusion with RRF, Memify content enrichment API.

• v0.7.0: Knowledge Graph + Session Dual-Layer Memory — entity graph storage, session-scoped memory isolation, dual-layer recall.

• v0.6.2: Security fix — access frequency weighting refinement, PBKDF2 260K→600K (OWASP 2023).

• v0.6.1: Architecture cleanup — capabilities decoupling, adapter optimization (PATCH).

• v0.6.0: Phase 3 Deprecated API Removal (Breaking Change) — legacy API cleanup, store_entry unified entry point.

• v0.5.4: Batch API — store_batch/delete_batch for bulk operations (Phase 2 optimization).

• v0.5.3: store_entry() core API — unified entry point (Phase 1 optimization).

• v0.5.2: Summary Layer + Progressive Disclosure — token-efficient prompt injection, auto-summarization.

• v0.5.1: Entity Normalization (Ontology-lite) — rule-based fuzzy entity matching.

• v0.5.0: Configurable access frequency weighting + selection access_boost.

• v0.4.0 (tech debt cleanup): PatternAnalyzer God Class split (1547→171 LOC facade + NoiseDetector + FeedbackDetector + MemoryPatternDetectors), mypy 536→0 errors with CI blocking gate, 394 public API docstrings added (50%→100%), 67 loose assertions strengthened (assertTrue→assertGreater), mypy.ini cleaned (python_version 3.10, removed unused sections). Maturity 74→80 (B-→B).

• v0.4.0: Quality Sprint — 134 unit tests for core Mixins, health_check MCP tool (28 total), TUI delete/edit, CLI rules subcommand grouping, audit SQLite persistence, removed @runtime_checkable, merged StorageAdapterProtocol, all Any types replaced (24→0), mypy+bandit CI

• v0.4.0: Protocol & Maturity Sprint — Mixin+Facade+Protocol 三层架构, 10个 Protocol 接口, 错误码体系 (CM-001~999), SQLite 连接池 (WAL+线程缓存), 加密升级 (PBKDF2 260K), E2E 测试补全 (+78), 监控框架 MVP, 插件系统 MVP, 权限系统 MVP, i18n 框架, 类型注解 ~82%, 72 new tests

• v0.3.0: Maturity & Architecture Sprint — God Class→8 Mixin, exception narrowing (173→15), TUI enhancement (+453 lines, Morandi palette), constants.py (28 named), lazy import cache, ghost feature audit, 71 new tests

• v0.2.5: Integration/E2E audit, ghost feature deprecation warnings, version chain validation, 83 new tests

• v0.2.4: Beta release — CI root fix, 24 security fixes, Glama TDQS boost, 6-gate CI pipeline

• v0.2.0: USB carry encryption, auto-backup, concurrent safety, PrefEval 83.0% (200 items), 8-client MCP setup

Contributing

git clone https://github.com/lulin70/carrymem.git
cd carrymem
pip install -e ".[dev]"
pytest

See Contributing Guide for details.

Citation

If you use CarryMem in your research, please cite:

@software{carrymem2026,
  title = {CarryMem: Persistent Memory for AI Agents with Preference Injection},
  author = {CarryMem Team},
  year = {2026},
  url = {https://github.com/carrymem/carrymem},
  note = {Preference injection accuracy 83.0\% measured by PrefEval protocol}
}

@inproceedings{chuang2025prefeval,
  title = {PrefEval: A Preference Evaluation Benchmark for LLMs},
  author = {Chuang, Yun-Nung and others},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year = {2025},
  note = {Oral presentation, Amazon Science}
}

Experimental Results (200 items, 10 inter-turns, Claude Sonnet 4)

Key insight: CarryMem achieves the highest accuracy while producing 24% fewer unhelpful responses than reminder-based approaches, demonstrating that proactive memory injection is more precise than full-context reminding.

License

MIT License — see LICENSE

CarryMem — Your AI finally remembers who you are. Only you own the data.