What is Agentmemory?

Agentmemory is a complete memory runtime designed specifically for coding agents like Hermes, Claude Code, and Codex. Instead of treating memory as an afterthought with vector stores or external databases, Agentmemory gives agents infinite, searchable memory that works out of the box. It captures every tool call, prompt, and session stop as compressed observations, then makes them recallable in milliseconds—all from a single Node process with zero external dependencies.

Who it's for

• AI coding agent users who want their agents to remember past sessions without hitting context limits or losing visibility into older interactions
• Developers building agentic workflows who need a production-ready memory layer that doesn't require setting up Redis, Kafka, or PostgreSQL
• Teams shipping CLI-based coding tools who want to add persistent, searchable memory to Claude Code, Codex, or OpenClaw with a single plugin install

Key features

Triple-stream recall

Agentmemory doesn't rely on a single retrieval method. It combines BM25 lexical search, vector similarity, and knowledge graph traversal, then reranks results on-device. The result is 95.2% R@5 on LongMemEval-S with P50 latency under 20 milliseconds on a laptop.

Auto-capture hooks

Twelve built-in hooks pipe every PreToolUse, PostToolUse, SessionStart, and Stop event directly into the memory pipeline. No glue code, no manual instrumentation—install the plugin and every agent action becomes a compressed observation automatically.

Autoconsolidation

Hourly sweeps compress raw observations into semantic memories, merge duplicates, and decay stale rows with retention scoring. Each delete emits a batched audit row, so you always know what was removed and why. This keeps memory lean without losing signal.

Native MCP and REST surface

Agentmemory exposes a full MCP server with tools like memory_save, memory_recall, memory_smart_search, and memory_sessions. Every MCP tool also has a REST twin under /agentmemory/*, so you can curl it, fetch it from a browser, or proxy it from your own agent.

What stands out

"At 1,000 observations, 80% of your built-in memories become invisible. Agentmemory keeps 100% searchable."

That's the core insight behind Agentmemory's design. Standard memory approaches lose visibility as context grows—observations get buried or dropped. Agentmemory's triple-stream retrieval and hourly consolidation ensure every observation stays searchable, regardless of volume. In benchmarked coding sessions, this translates to up to 95% fewer tokens per session and 200x more tool calls before hitting context limits.

Worth checking out if…

You're tired of your coding agent forgetting what it did two sessions ago, or you're tired of wiring up Redis, Kafka, and a vector store just to give an agent persistent memory. Agentmemory runs as a single Node process, stores state as JSON on disk, and ships with a real-time viewer on port 3113. It's 100% open source, trending on GitHub with 5,000+ stars, and works with Claude Code, Codex, Hermes, OpenClaw, and Pi out of the box.