engram
Your AI coding assistant forgets everything.
Persistent memory that learns from every session.
$ npm install -g engramx $ cd my-project && engram init $ engram install-hook → hook installed · 451 nodes, 1005 edges · 0 tokens used ✓ Your AI now has persistent memory. Structure over raw bytes.
Your code has structure that compiles away — calls, imports, inherits, throws, documents. engram reads that structure and stores it as a knowledge graph. When an AI agent asks for a file, engram delivers the relevant slice of graph instead — 10× to 150× fewer tokens than full file content, and more informative.
Hook-based interception
Sentinel installs a PreToolUse hook for Read / Edit / Write / Bash. Every access flows through engram first. The IDE asks for bytes; it receives knowledge.
6 context providers
structure— AST graphmistakes— regret buffergit— co-change pairsmempalace— semantic memorycontext7— library docsobsidian— notes, ADRs
88.1% token reduction
Aggregate across 8 real coding tasks. Cache hit rate 99% in the hot path. 673 tests passing on CI across Ubuntu + Windows × Node 20 / 22.
Index once. Intercept always.
- 01once per project
Build the graph
`engram init` walks your repo, extracts nodes and edges from 10 languages via tree-sitter, writes to .engram/graph.db. Incremental on re-runs.
- 02once per machine
Install Sentinel
`engram install-hook` registers a PreToolUse hook globally. From that moment every Claude Code file access flows through engram first.
- 03every session
Query structure
Use `mcp-engram query "question"` or let Sentinel inject context on every Read. Past mistakes surface at top with ⚠️ warning.
- 04across sessions
Memory compounds
Session learnings, resolved mistakes, co-change pairs accumulate in the graph. The agent keeps getting smarter as you work.
Cirvgreen builds two product lines. Open-source dev tools (engram, Apache 2.0) and premium WordPress plugins (Cirv Box, Guard, Pulse, Comply). Both share an engineering discipline — do one thing, measure honestly, stay local-first.