ai-memory

Product Introduction

1. Definition: AI Memory is an open-source, server-based long-term memory system designed for AI coding agents and command-line interfaces (CLIs). It functions as a persistent, context-aware knowledge base that operates via the Model Context Protocol (MCP) and lifecycle hooks.
2. Core Value Proposition: It solves the critical problem of context loss between AI agent sessions, enabling seamless handoffs and continuity across different AI coding tools like Claude Code, GitHub Copilot, Cursor, and Devin CLI. Its primary value is providing persistent, shareable memory to maintain development workflow efficiency.

Main Features

1. Zero-Friction Lifecycle Capture: The system automatically captures sanitized observations from AI agent sessions via installed hooks for events like SessionStart, ToolUse, and SessionEnd. It enforces configurable capture exclusions (e.g., for node_modules) and stores bounded observations (e.g., 16 KiB for prompts, 2 KB for tool excerpts) in a local spool before server processing, requiring no manual intervention from the developer.
2. Git-Versioned Markdown Wiki: All consolidated memory is stored as plain Markdown files within a Git repository. This structure allows for full-text search via SQLite's FTS5, easy browsing with tools like Obsidian, and simple backup/versioning using standard Git commands or rsync. Pages are compiled from observations at session boundaries, not raw logs, creating a coherent, versioned knowledge base.
3. Managed Workstreams & Native Session Resume: The ai-memory run command wrapper enables transparent continuity across different AI harnesses (e.g., Claude Code to Codex). It manages a portable ledger of visible events and can resume the native session of the target agent, providing higher-fidelity context handoff than a simple summary packet alone.
4. Multi-Agent & Multi-Client Support: It provides first-party integration for over 15 AI coding agents and CLIs, including Claude Code, Cursor, OpenCode, Devin CLI, and GitHub Copilot Agent Mode. Support includes generating agent-specific MCP configuration files, lifecycle hook bundles, and in some cases, native plugin extensions (e.g., for Pi editor).
5. Entity-Assisted & Authority-Aware Recall: The system extracts canonical entities (nouns) from pages to aid lexical recall. Its retrieval pipeline combines FTS5, entity-match RRF, graph-neighbor RRF, and optional vector RRF, then applies a bounded authority adjustment to favor maintained pages in _rules/ or decisions/ namespaces over episodic evidence, without filtering them out entirely.

Problems Solved

1. Pain Point: AI coding agents suffer from complete amnesia between sessions, forcing developers to repeatedly re-explain project architecture, past decisions, failed approaches, and open questions when switching tools or restarting work.
2. Target Audience: The primary users are software developers, engineering teams, and tech leads who utilize multiple AI-powered coding assistants (multi-agent workflows) and require persistent context to avoid productivity loss. It is especially valuable for consultants, open-source maintainers, and teams working in complex, long-lived codebases.
3. Use Cases: Essential scenarios include: quitting Claude Code mid-task and resuming hours later in Codex with full context; onboarding a new team member's AI agent to a project's historical rationale; maintaining a searchable decision log for a legacy system; and ensuring continuity when alternating between GUI-based agents (Cursor) and CLI-based agents (Devin CLI).

Unique Advantages

1. Differentiation: Unlike simple note-taking plugins or vector database solutions, AI Memory is a structured, event-driven system that automatically builds context from agent behavior, not just manual write_note commands. Compared to proprietary agent memory features, it is vendor-agnostic, open-source, and stores data in a transparent, portable Markdown+Git format.
2. Key Innovation: Its hybrid retrieval system that prioritizes "source-of-truth" pages (like _rules/) through algorithmic authority signals, while still allowing historical session evidence to be found. This mimics a team's trust in documented procedures over chat history, a nuance often missing in pure vector similarity search.

Frequently Asked Questions (FAQ)

1. How does AI Memory handle sensitive data like API keys? AI Memory's capture hooks are designed to exclude sensitive tool outputs by default and can be configured with a [capture] ignore_paths policy in a .ai-memory.toml file. It also integrates with secret detection tools via a .gitleaksignore file, and the server never transmits captured data to external LLM providers without explicit user configuration.
2. Can AI Memory be used in a team or multi-user environment? Yes, the server supports multi-user operation. When configured with [slots] per_user = true, context written to the _slots/ namespace is isolated per authenticated user, while the core wiki remains a shared, project-wide resource. Administrator endpoints like /admin/activity/by-client provide visibility into usage.
3. What happens if the AI Memory server is offline? Lifecycle hooks write observations to a local disk spool (.ai-memory/spool/). When the server is restarted, it processes this backlog, ensuring no context is lost due to temporary downtime. The CLI commands will fail if they cannot reach the server.
4. Does using AI Memory slow down my AI coding agent? The hook system is designed to be fire-and-forget and asynchronous. Observations are spooled locally with minimal overhead. The actual processing (consolidation, embedding) happens on the server side, not blocking the agent's primary interaction loop.
5. How do I search my AI Memory wiki? You can use the built-in /web browser UI for full-text search and navigation, use the ai-memory query CLI command, or directly grep the Markdown files in the wiki repository. The system uses SQLite FTS5 for efficient project-scoped search.