agentmemory is an open-source persistent memory layer purpose-built for AI coding agents like Claude Code, Cursor, Cline, Gemini CLI, Hermes, and OpenClaw. Released by rohitg00 and trending in May 2026 with the v0.9.9 release, the project addresses a problem every developer using AI coding tools has hit: the agent forgets your project the moment a session ends, and you spend the first ten minutes of every new session re-explaining the codebase, the conventions, the gotchas. agentmemory captures that context once and feeds the relevant slice back automatically.

Numbers Worth Quoting

The README reports 95.2% retrieval R@5 and 92% fewer tokens compared to pasting full context into a new session. These are benchmark-driven numbers rather than marketing claims and are backed by a public retrieval evaluation in the repo. The token reduction is the more interesting figure in practice, because for any team paying for AI coding tools, replacing a 30k-token context paste with a few hundred tokens of relevant retrieved memory is a direct cost reduction.

Automatic Capture via 12 Hooks

Most memory tools in this space require manual tagging or explicit save commands, which means users forget to use them. agentmemory ships 12 hooks that capture observations automatically as the agent works: file edits, command outputs, test results, conversation turns, and tool calls all flow into memory without manual configuration. The default behavior is zero-configuration capture, with privacy filtering and deduplication applied before storage.

Multi-Stream Retrieval

Retrieval combines three streams: BM25 keyword matching for exact-name queries ("that function in payments.ts"), vector embeddings for semantic similarity, and a knowledge graph that captures relationships between code entities, tasks, and decisions. The combination handles both lookup-style queries and broader "what did we decide about X" questions that pure vector search tends to miss.

4-Tier Memory Consolidation

The system implements a four-tier consolidation pipeline inspired by cognitive memory research: working memory holds the active session, episodic memory stores specific events, semantic memory distills repeated patterns into general facts, and procedural memory captures how-to knowledge such as build commands and deployment steps. Older memories get compressed or forgotten according to use frequency, which keeps the store from bloating over months of work.

51 MCP Tools and 107 REST Endpoints

Integration is broad: 51 Model Context Protocol tools make the memory store accessible to any MCP-aware agent, and a 107-endpoint REST API covers non-MCP integrations. A real-time viewer on port 3113 lets developers watch memories form and inspect what is being retrieved for any given query, which is useful both for debugging and for trust.

Built on the iii Engine

Unusually, agentmemory does not lean on the traditional Express + Postgres + Redis stack. It is built on the iii engine, a primitives-based runtime that replaces conventional infrastructure with composable workers extended via iii worker add. Embeddings can run locally with no API key, and LLM-powered compression is optional. This makes the project notably easier to self-host than memory tools that assume a cloud database.

Where It Fits

agentmemory is a strong fit for any developer or team using a coding agent daily. It is not a replacement for the agent itself, but a sidecar that makes the agent dramatically better at remembering the project. For 2026, where most teams have settled on one or two AI coding tools, a vendor-neutral memory layer that works across all of them is exactly the kind of infrastructure the space has been missing.