Engram: give your coding agents a memory (built on FalkorDB)

Coding agents spend a third of their tool calls just searching. Engram turns your repo into a FalkorDB graph (code, git history and what your agents already did) and hands the agent a ~200-token map of where to look before its first tool call. Measured, not claimed: • pallets/click, 144 Claude Code runs, same model: −25% input tokens, −29% tool calls, −23% time • networkx (128k lines): −24% tokens, −40% tool calls • yt-dlp (248k lines), 30 real GitHub issues: right file in the top 3 for 22/30, zero LLM tokens • Live on 4 of those issues: −31% tokens, −33% tool calls, same right answer 4/4 Install (no Docker, no API keys, no LLM calls of its own): curl -fsSL https://raw.githubusercontent.com/ManojRakshu/engram/main/install.sh | sh Code, benchmark harness and raw results: https://github.com/ManojRakshu/engram Built for Graph Hacks: Context for AI Agents (WeMakeDevs × FalkorDB). 0:00 The problem: agents keep searching 0:23 What Engram does 0:37 Why FalkorDB: a graph walk, not a grep 0:57 Benchmark: 144 runs 1:13 A real, messy project: yt-dlp 1:34 Memory, blast radius, and install