Engram v0.2 ships local knowledge graph
Engram is a persistent codebase memory tool for AI coding assistants that keeps its graph in a local SQLite file at `.engram/graph.db` and avoids cloud services, telemetry, embeddings, native dependencies, and LLM calls during indexing. The v0.2 release adds skill-directory indexing, task-aware graph generation for bug-fix/feature/refactor workflows, a regret buffer that surfaces past mistakes, a new `list_mistakes` MCP tool, and several correctness fixes around JSON-RPC framing and atomic initialization. It is positioned for people running local models who need smaller, more structured context than dumping raw files into the prompt.
Strong release if the claims hold up in practice: this is opinionated infrastructure for local-first coding agents, not another generic vector-store wrapper.
- –The local-only angle is the real differentiator: deterministic extraction, plain SQLite, and no runtime network calls is a clearer trust story than “local-ish” tools that still depend on cloud APIs.
- –The task-aware generation is practical for small local models; trimming context around hot files, architecture, or known mistakes is more useful than yet another search box.
- –Indexing `~/.claude/skills/` is a smart extension because it links codebase context with agent-specific workflows instead of treating them separately.
- –The biggest risk is accuracy and recall: regex-based mining is fast and portable, but it will miss structure that a parser or richer extractor would catch.
- –This is most compelling for solo devs and small teams already leaning into Claude Code, Cursor, or other agentic workflows on constrained hardware.
DISCOVERED
1d ago
2026-04-10
PUBLISHED
2d ago
2026-04-10
RELEVANCE
AUTHOR
SearchFlashy9801