// Agent profile
ai-memory (by akitaonrails / Fabio Akita) is a Rust-compiled single-binary MCP/HTTP server that gives coding agents persistent long-term memory with structured cross-vendor handoff. Instead of dumping raw transcripts, it compiles observations into coherent markdown wiki pages backed by SQLite with FTS5 full-text search, optional vector embeddings, and git-versioned storage for time-travel recovery. The handoff protocol captures from-agent, to-agent, context, open questions, files touched, and model used — enabling mid-task switches between Claude Code, Codex, Cursor, Gemini CLI, and 12 other harnesses without losing context. Zero external service dependencies for core functionality; embeddings are optional. MIT licensed.
ai-memory solves a problem that every multi-agent operator hits: context dies when you switch tools. Move from Claude Code to Cursor to Codex CLI and each session starts cold, unaware of what the previous agent learned about your codebase, your preferences, or the bug you were halfway through diagnosing.
The project takes a different architectural approach from most memory layers. Rather than storing and retrieving raw conversation logs, ai-memory compiles observations into structured markdown wiki pages — a knowledge base that grows with each session. The storage layer is SQLite with FTS5 full-text search, git-versioned so every state is recoverable. Vector embeddings are optional, not required — the core search stack works with zero external API cost.
Structured handoff protocol: The defining feature. When switching agents, ai-memory captures a structured handoff record: which agent is handing off, which is receiving, the current context and open questions, files touched, and the model used. The receiving agent gets a coherent briefing instead of starting from zero.
15+ CLI integrations: Claude Code, OpenAI Codex, Cursor, Gemini CLI, Grok Build CLI, Kimi Code, Kiro CLI, Antigravity CLI, Devin CLI, OpenCode, Command Code, OMP, Pi, Crush, OpenClaw, Zero, and Hermes (community plugin). The cross-harness breadth is the value proposition — a memory layer locked to one vendor is structurally narrower.
Wiki compilation: Observations are compiled into coherent markdown pages organized by project and topic. This is meaningfully different from transcript retrieval — the wiki reads like documentation, not like searching through chat logs.
Retrieval stack: FTS5 full-text search with entity matching and graph-neighbor reciprocal rank fusion. Optional vector RRF when embeddings are configured. The architecture decision to make vectors optional keeps the core dependency-free.
Multi-agent operators who switch between coding CLIs within a single project. Teams where different developers use different agent tools but need shared project memory. Long-running projects where session context exceeds what fits in a CLAUDE.md or .cursorrules file. Operators who want recoverable, auditable memory with git versioning.
The Rust binary compiles cleanly but requires either Docker, a native build via Cargo, or a platform-specific binary. Optional embedding features require API keys for an LLM provider (Anthropic, OpenAI, Gemini), adding cost for semantic search — but core FTS5 search is free. The wiki compilation step uses LLM calls when configured, which means memory quality scales with the consolidation model’s capability.
Developers and teams running multiple coding agent CLIs who need context to survive tool switches. Operators who want their agent memory versioned, searchable, and auditable rather than ephemeral. Anyone who has lost context switching from one agent to another mid-task and wants a structured solution.
Builder.io's open-source framework for building agent-native applications — shared actions, SQL-backed state, identity, tools, skills, jobs, observability, and UI surfaces that all work together.
CLI orchestration tool that installs and configures platform connectors for 16+ social and web sources — Twitter/X, Reddit, YouTube, GitHub, Bilibili, and more. Zero API fees.
Persistent memory layer for AI coding agents — benchmark-backed (95.2% on LongMemEval-S), 92% fewer tokens per session vs full-context pasting, zero manual memory.add() calls.