Aivectormemory

🌐 [简体中文](docs/README.zh-CN.md) | [繁體中文](docs/README.zh-TW.md) | English | [Español](docs/README.es.md) | [Deutsch](docs/README.de.md) | [Français](docs/README.fr.md) | [日本語](docs/README.ja.md)

AIVectorMemory

Give your AI coding assistant a memory — Cross-session persistent memory MCP Server

---

> Still using CLAUDE.md / MEMORY.md as memory? This Markdown-file memory approach has fatal flaws: the file keeps growing, injecting everything into every session and burning massive tokens; content only supports keyword matching — search "database timeout" and you won't find "MySQL connection pool pitfall"; sharing one file across projects causes cross-contamination; there's no task tracking, so dev progress lives entirely in your head; not to mention the 200-line truncation, manual maintenance, and inability to deduplicate or merge. > > AIVectorMemory is a fundamentally different approach. Local vector database storage with semantic search for precise recall (matches even when wording differs), on-demand retrieval that loads only relevant memories (token usage drops 50%+), automatic multi-project isolation with zero interference, and built-in issue tracking + task management that lets AI fully automate your dev workflow. All data is permanently stored on your machine — zero cloud dependency, never lost when switching sessions or IDEs.

✨ Core Features

| Feature | Description | |---------|-------------| | 🧠 Cross-Session Memory | Your AI finally remembers your project — pitfalls, decisions, conventions all persist across sessions | | 🔍 Hybrid Smart Search | FTS5 full-text + vector semantic dual-path search, RRF fusion ranking + composite scoring (recency × frequency × importance), far more precise than pure vector search | | 🐛 Issue Tracking | Built-in Issue Tracker — discover → investigate → fix → archive, full lifecycle. AI manages bugs automatically | | 📋 Task Management | Spec → task breakdown → nested subtasks → status sync → linked archival. AI drives the complete dev workflow | | 🚦 Session State | Blocking management + breakpoint resume + progress tracking, seamless handoff across sessions and context compaction | | 🪝 Hooks + Steering | Auto-inject workflow rules + behavior guard hooks, consistent AI behavior guaranteed — no need to repeat instructions | | 🧬 Memory Evolution | Contradiction detection auto-supersedes stale knowledge + short-term → long-term auto-promotion + 90-day auto-archive, self-evolving memory | | 📊 Desktop App + Web Dashboard | Native desktop app (macOS/Windows/Linux) + Web dashboard, 3D vector network reveals knowledge connections at a glance | | 💰 Save 50%+ Tokens | Stop copy-pasting project context every conversation. Semantic retrieval on demand, no more bulk injection | | 🏠 Fully Local | Zero cloud dependency. ONNX local inference, no API Key, data never leaves your machine | | 🔌 11 IDEs Covered | Cursor / Kiro / Claude Code / Windsurf / VSCode / Copilot / OpenCode / Trae / Codex / Antigravity / OpenClaw — one-click install & uninstall | | 📁 Multi-Project Isolation | One DB for all projects, auto-isolated with zero interference, seamless project switching | | 🔄 Smart Dedup | Similarity > 0.95 auto-merges updates, keeping your memory store clean — never gets messy over time | | 🌐 7 Languages | 简体中文 / 繁體中文 / English / Español / Deutsch / Français / 日本語, full-stack i18n for dashboard + Steering rules |

QQ群：1085682431  |  微信：changhuibiz 共同参与项目开发加QQ群或微信交流

Login

Project Selection

Overview & Vector Network

🏗️ Architecture

┌─────────────────────────────────────────────────┐
│                   AI IDE                         │
│  OpenCode / Codex / Claude Code / Cursor / ...  │
└──────────────────────┬──────────────────────────┘
                       │ MCP Protocol (stdio)
┌──────────────────────▼──────────────────────────┐
│              AIVectorMemory Server               │
│                                                  │
│  ┌──────────┐ ┌──────────┐ ┌──────────────────┐ │
│  │ remember │ │  recall   │ │   auto_save      │ │
│  │ forget   │ │  task     │ │   status/track   │ │
│  └────┬─────┘ └────┬─────┘ └───────┬──────────┘ │
│       │            │               │             │
│  ┌────▼────────────▼───────────────▼──────────┐  │
│  │         Embedding Engine (ONNX)            │  │
│  │      intfloat/multilingual-e5-small        │  │
│  └────────────────────┬───────────────────────┘  │
│                       │                          │
│  ┌────────────────────▼───────────────────────┐  │
│  │     SQLite + sqlite-vec (Vector Index)     │  │
│  │     ~/.aivectormemory/memory.db            │  │
│  └────────────────────────────────────────────┘  │
└──────────────────────────────────────────────────┘

🚀 Quick Start

Option 1: pip install (Recommended)

# Install
pip install aivectormemory

# Upgrade to latest version
pip install --upgrade aivectormemory

# Navigate to your project directory, one-click IDE setup
cd /path/to/your/project
run install

run install interactively guides you to select your IDE, auto-generating MCP config, Steering rules, and Hooks — no manual setup needed.

> macOS users note: > - If you get externally-managed-environment error, add --break-system-packages > - If you get enable_load_extension error, your Python doesn't support SQLite extension loading (macOS built-in Python and python.org installers don't support it). Use Homebrew Python instead: > ``bash > brew install python > /opt/homebrew/bin/python3 -m pip install aivectormemory > ``

Option 2: uvx (zero install)

No pip install needed, run directly:

cd /path/to/your/project
uvx aivectormemory install

> Requires uv to be installed. uvx auto-downloads and runs the package — no manual installation needed.

Option 3: Manual configuration

{
  "mcpServers": {
    "aivectormemory": {
      "command": "run",
      "args": ["--project-dir", "/path/to/your/project"]
    }
  }
}

📍 IDE Configuration File Locations

| IDE | Config Path | |-----|------------| | Kiro | .kiro/settings/mcp.json | | Cursor | .cursor/mcp.json | | Claude Code | .mcp.json | | Windsurf | .windsurf/mcp.json | | VSCode | .vscode/mcp.json | | Trae | .trae/mcp.json | | OpenCode | opencode.json | | Codex | .codex/config.toml |

For Codex, use project-scoped TOML instead of JSON:

[mcp_servers.aivectormemory]
command = "run"
args = ["--project-dir", "/path/to/your/project"]

> Codex only loads project-scoped .codex/config.toml after the repository is marked as a trusted project.

🛠️ 9 MCP Tools

remember — Store a memory

content (string, required)   Memory content in Markdown format
tags    (string[], required)  Tags, e.g. ["pitfall", "python"]
scope   (string)              "project" (default) / "user" (cross-project)

Similarity > 0.95 auto-updates existing memory, no duplicates.

recall — Semantic search

query   (string)     Semantic search keywords
tags    (string[])   Exact tag filter
scope   (string)     "project" / "user" / "all"
top_k   (integer)    Number of results, default 5

Vector similarity matching — finds related memories even with different wording.

forget — Delete memories

memory_id  (string)     Single ID
memory_ids (string[])   Batch IDs

status — Session state

state (object, optional)   Omit to read, pass to update
  is_blocked, block_reason, current_task,
  next_step, progress[], recent_changes[], pending[]

Maintains work progress across sessions, auto-restores context in new sessions.

track — Issue tracking

action   (string)   "create" / "update" / "archive" / "list"
title    (string)   Issue title
issue_id (integer)  Issue ID
status   (string)   "pending" / "in_progress" / "completed"
content  (string)   Investigation content

task — Task management

action     (string, required)  "batch_create" / "update" / "list" / "delete" / "archive"
feature_id (string)            Linked feature identifier (required for list)
tasks      (array)             Task list (batch_create, supports subtasks)
task_id    (integer)           Task ID (update)
status     (string)            "pending" / "in_progress" / "completed" / "skipped"

Links to spec docs via feature_id. Update auto-syncs tasks.md checkboxes and linked issue status.

readme — README generation

action   (string)    "generate" (default) / "diff" (compare differences)
lang     (string)    Language: en / zh-TW / ja / de / fr / es
sections (string[])  Specify sections: header / tools / deps

Auto-generates README content from TOOL_DEFINITIONS / pyproject.toml, multi-language support.

auto_save — Auto save preferences

preferences  (string[])  User-expressed technical preferences (fixed scope=user, cross-project)
extra_tags   (string[])  Additional tags

Auto-extracts and stores user preferences at end of each conversation, smart dedup.

graph — Code knowledge graph

action       (string, required)  "query" / "trace" / "batch" / "add_node" / "add_edge" / "remove" / "refresh"
name         (string)            Entity name (add_node/query)
entity_type  (string)            Entity type: function/class/module/api/table/config (add_node/query)
file_path    (string)            File path, auto-converts to relative (add_node/query/refresh)
source       (string)            Source node name or ID (add_edge)
target       (string)            Target node name or ID (add_edge)
relation     (string)            Relation type: calls/imports/inherits/uses/depends_on/contains (add_edge/trace)
start        (string)            Start node name or ID (trace)
direction    (string)            Traversal direction: "up" / "down" / "both" (trace)
max_depth    (integer)           Max traversal depth, default 3 (trace)

Manages function call chains, data flows, and dependency relationships. Trace upstream/downstream impact before code changes.

📊 Web Dashboard

run web --port 9080
run web --port 9080 --quiet          # Suppress request logs
run web --port 9080 --quiet --daemon  # Run in background (macOS/Linux)

Visit http://localhost:9080 in your browser. Default username admin, password admin123 (can be changed in settings after first login).

• Multi-project switching, memory browse/search/edit/delete/export/import
• Semantic search (vector similarity matching)
• One-click project data deletion
• Session status, issue tracking
• Tag management (rename, merge, batch delete)
• Token authentication protection
• 3D vector memory network visualization
• 🌐 Multi-language support (简体中文 / 繁體中文 / English / Español / Deutsch / Français / 日本語)

    

Scan to join WeChat group  |  Scan to join QQ group

⚡ Pairing with Steering Rules

AIVectorMemory is the storage layer. Use Steering rules to tell AI when and how to call these tools.

Running run install auto-generates Steering rules and Hooks config — no manual setup needed.

| IDE | Steering Location | Hooks | |-----|------------------|-------| | Kiro | .kiro/steering/aivectormemory.md | .kiro/hooks/*.hook | | Cursor | .cursor/rules/aivectormemory.md | .cursor/hooks.json | | Claude Code | CLAUDE.md (appended) | .claude/settings.json | | Windsurf | .windsurf/rules/aivectormemory.md | .windsurf/hooks.json | | VSCode | .github/copilot-instructions.md (appended) | .claude/settings.json | | Trae | .trae/rules/aivectormemory.md | — | | OpenCode | AGENTS.md (appended) | .opencode/plugins/*.js | | Codex | AGENTS.md (appended) | — |

📋 Steering Rules Example (auto-generated)

# AIVectorMemory - Workflow Rules

## 1. New Session Startup (execute in order)

1. `recall` (tags: ["project-knowledge"], scope: "project", top_k: 100) load project knowledge
2. `recall` (tags: ["preference"], scope: "user", top_k: 20) load user preferences
3. `status` (no state param) read session state
4. Blocked → report and wait; Not blocked → enter processing flow

## 2. Message Processing Flow

- Step A: `status` read state, wait if blocked
- Step B: Classify message type (chat/correction/preference/code issue)
- Step C: `track create` record issue
- Step D: Investigate (`recall` pitfalls + read code + find root cause)
- Step E: Present plan to user, set blocked awaiting confirmation
- Step F: Modify code (`recall` pitfalls before changes)
- Step G: Run tests to verify
- Step H: Set blocked awaiting user verification
- Step I: User confirms → `track archive` + clear block

## 3. Blocking Rules

Must `status({ is_blocked: true })` when proposing plans or awaiting verification.
Only clear after explicit user confirmation. Never self-clear.

## 4-9. Issue Tracking / Code Checks / Spec Task Mgmt / Memory Quality / Tool Reference / Dev Standards

(Full rules auto-generated by `run install`)

🔗 Hooks Config Example (Kiro only, auto-generated)

Auto-save on session end removed. Dev workflow check (.kiro/hooks/dev-workflow-check.kiro.hook):

{
  "enabled": true,
  "name": "Dev Workflow Check",
  "version": "1",
  "when": { "type": "promptSubmit" },
  "then": {
    "type": "askAgent",
    "prompt": "Core principles: verify before acting, no blind testing, only mark done after tests pass"
  }
}

🇨🇳 Users in China

The embedding model (~200MB) is auto-downloaded on first run. If slow:

export HF_ENDPOINT=https://hf-mirror.com

Or add env to MCP config:

{
  "env": { "HF_ENDPOINT": "https://hf-mirror.com" }
}

📦 Tech Stack

| Component | Technology | |-----------|-----------| | Runtime | Python >= 3.10 | | Vector DB | SQLite + sqlite-vec | | Embedding | ONNX Runtime + intfloat/multilingual-e5-small | | Tokenizer | HuggingFace Tokenizers | | Protocol | Model Context Protocol (MCP) | | Web | Native HTTPServer + Vanilla JS |

📋 Changelog

v2.4.5

Patch: Hard Constraints Against Opus 4.7 Default Tendencies

• 🚫 §1 added No Clarification-Style Follow-ups: forbid re-asking "phased or one-shot / full or partial / should I do X / do A first or B first" for imperative commands; under ambiguity, execute the most complete scope
• 🚫 §1 added No Defensive Reporting: forbid wording like "kept per instruction / marked pending / non-critical path / unnecessary sub-tests / for later iteration" as excuse for unexecuted items
• 📋 §1 added Report Format: forbid Phase A/B/C/D list + "Final Status" + "Not Done (Kept Per Instruction)" three-section format; when user says "do all", no "Not Done/Kept" section allowed
• 🎯 Root cause: counteract Opus 4.7's default "defensive reporting", "clarification follow-ups", and "structured checklist" tendencies compared to 4.6
• 🔄 7-language rule files (STEERINGCONTENT + DEVWORKFLOW_PROMPT) fully synced with CLAUDE.md v2.4.5 updates

v2.4.4

Patch: Full A-I Message Processing Flow Alignment

• 🧩 CLAUDE.md §4 message processing B routes fully expanded: all 4 branches (casual/correction/preference/other) unified to terminate at I(user confirm & archive), eliminating incomplete "stop at F" flows
• ⚙️ inject-workflow-rules.sh message type judgment section fully aligned with §4 B: 4 routes with consistent granularity
• 🔧 Fixed 3 conflicts: route granularity inconsistency (2 vs 4 routes) / B and E responsibility mixing ("solution+block" misplaced) / G/H/I flow missing
• 📝 Unified violation clause: "Proceeding to C/D/E/F steps without outputting judgment result = violation"
• 🔄 7-language rule files (STEERINGCONTENT + DEVWORKFLOW_PROMPT) fully synced with CLAUDE.md v2.4.4 updates

v2.4.3

Patch: Rule Enforcement & Graph Visualization

• 🧠 §4.B: Two-step mandatory structure (understand message → determine type)

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Edlineas
• Source: Edlineas/aivectormemory
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.