agentmemory is Rohit Ghumare's open-source persistent-memory layer for AI coding agents, designed to end the recurring frustration of re-explaining your codebase, your conventions, and your past decisions every time a session starts. With 17,879 stars, 1,462 forks, and an Apache 2.0 license, the project bills itself as the number-one persistent memory layer for Claude Code, Cursor, Gemini CLI, Codex CLI, Hermes, OpenClaw, pi, OpenCode, and any other MCP-compatible client. It is built on the iii engine and extends Andrej Karpathy's LLM Wiki pattern with confidence scoring, lifecycle management, knowledge graphs, and hybrid search — a design that first went viral as a GitHub gist with 1,200 stars and 172 forks before becoming a full implementation.

A Single Memory Server Across Every Agent

The architectural choice that makes agentmemory distinctive is that it runs as a single local memory server on port 3111 that every agent talks to through hooks, MCP, or REST. Install once with npm install -g @agentmemory/agentmemory, then connect Claude Code, Codex, Cursor, Gemini CLI, Hermes, OpenClaw, OpenCode, and any other MCP client to the same store. This means a fact captured during a Claude Code session is available in a later Cursor session, and a Codex-side decision is visible to a Gemini CLI agent, without manual export or re-prompting. There are zero external database dependencies — the server runs entirely on local storage.

Benchmarked Recall and Token Savings

The project publishes concrete benchmark numbers rather than vague claims of better memory: 95.2 percent retrieval recall at R@5 on the project's internal evaluation, 92 percent fewer tokens consumed compared to dumping full context into the prompt, and a regression suite of more than 950 passing tests. The 92-percent token reduction is the practically important number for anyone running on metered API access, because it directly translates to a 10x reduction in per-session cost for memory-dependent workflows.

Hybrid Search, Confidence Scoring, and Knowledge Graphs

Under the hood agentmemory is more than a vector store. It combines hybrid search across semantic and keyword indexes, a confidence-scoring mechanism that lets the agent decide how much to trust a recalled fact, lifecycle management so stale or contradicted memories decay rather than accumulate forever, and a knowledge-graph layer that captures relationships between entities the agent has seen. This is closer to a structured cognition layer than a simple recall tool, and it is the reason the project advertises 53 distinct MCP tools rather than the usual handful.

Twelve Auto-Hooks for Claude Code, Six for Codex

Integration depth varies by harness and agentmemory's documentation is unusually honest about that. Claude Code gets a native plugin with 12 automatic hooks plus MCP, so memory capture and recall happen without the user thinking about it. Codex CLI gets 6 hooks plus MCP. OpenClaw, Hermes, Cursor, Gemini CLI, OpenCode, and pi get native plugins plus MCP coverage, while any other MCP client can use the REST API directly. The 53-tool MCP surface is what makes the deep integrations possible.

Pragmatic Distribution

The project ships through npm both as a global install (npm install -g @agentmemory/agentmemory) and as a one-shot npx (npx @agentmemory/agentmemory), with explicit warnings about npx caching pitfalls and an inline upgrade prompt the first time it runs. The first-run developer experience includes a demo command that seeds sample sessions and proves recall is working, plus an agentmemory connect <agent> helper that wires the chosen harness in one step. There is also a real-time viewer for inspecting what the memory server has captured, which removes one of the biggest objections to agent-memory systems — that they are black boxes the user cannot audit.

Why It Matters

Persistent memory is the missing primitive in 2026 agent workflows. Sessions remain stateless by default and the workaround — pasting a giant CLAUDE.md or AGENTS.md prefix into every conversation — does not scale across team members, machines, or harnesses. agentmemory replaces that workaround with a benchmarked, audit-able, harness-agnostic memory server that takes the cost of remembering off the user and puts it where it belongs: in the toolchain.