Agentmemory ships benchmarked memory layer for copilots
agentmemory is an open-source persistent memory runtime for AI coding agents that captures prompts, tool calls, and session events, then consolidates them into searchable memory across Claude Code, Cursor, Gemini CLI, Codex CLI, OpenCode, and other MCP or HTTP clients. The project positions itself as a full memory engine rather than a library or vector store, with auto-capture hooks, hybrid retrieval over BM25 plus vector plus graph search, and a live viewer for replaying sessions. Its README highlights strong retrieval and token-efficiency claims, including 95.2% R@5 on LongMemEval-S and roughly 92% fewer tokens than pushing full context.
Hot take: this is trying to become the default memory substrate for coding agents, not just another “remember my notes” addon.
- –The benchmark angle matters: 95.2% R@5 and the token-savings story make the product legible to serious builders, not just hobbyists.
- –Its biggest advantage is portability: MCP, REST, hooks, and support for multiple agents reduce lock-in.
- –The architecture is opinionated but practical: auto-capture, consolidation, replay, and graph-aware recall are all aimed at eliminating manual memory curation.
- –The main risk is complexity; a memory system that deep may be powerful, but it also raises the bar for setup, maintenance, and trust.
DISCOVERED
1h ago
2026-05-09
PUBLISHED
1h ago
2026-05-09
RELEVANCE