May 9, 2026AgentsToolMCP

agentmemory Gives AI Coding Agents Real Memory

Claude Code, Cursor, Gemini CLI, all the rest of them — they're amnesiacs. Every new session you start by re-explaining the codebase, re-pasting the conventions, re-telling them what failed last time. agentmemory just shipped v0.9.5 and the entire pitch is killing that loop.

Sit a hook on your coding agent. Every tool call gets captured automatically. Decisions, errors, fixes, conventions — all of it goes into a four-tier memory system inspired by how human memory consolidates. When the next session starts, the agent reaches in and pulls what's relevant via BM25 plus vector search. No copy-paste, no "here's the context again."

The numbers are why this matters. 95.2% retrieval accuracy at rank-5 on LongMemEval-S, 92% token reduction versus dumping full context. The author claims about $10/year per developer instead of $500. That's not a 10% improvement, that's a different cost regime — the kind of gap that makes "don't bother with memory" stop being a defensible default.

Built on iii engine, 21,800 lines of TypeScript, 800 tests, 51 MCP tools. Real-time viewer at port 3113 lets you actually inspect what the agent remembers, which is the part most memory products skip. v0.9.5 specifically pins iii-engine to v0.11.2 (architectural changes upstream), fixes BM25 indexing on saved memories, and adds vector index corruption guards.

github.com/rohitg00/agentmemory — 3.4K stars, +518 today. The category of "long-term memory layer for coding agents" was wide open six months ago. It is closing fast.
← Previous
GitHub Stars Daily Spotlight — May 10, 2026
Next →
Nvidia Has Already Spent $40B Buying AI Equity This Year
← Back to all articles

Comments

Loading...
>_