Engram

Caches your codebase structure in local SQLite so AI agents stop re-reading the same files on every request. Exposes MCP tools for querying the indexed graph, tracking mistakes from git revert history, and surfacing past corrections before the agent repeats them. The bi-temporal mistake table auto-populates from your repo's revert commits and fires guards when an agent touches files that match known failure patterns. Measured 89.1% token reduction on an 87-file benchmark. Works across Claude Code, Cursor, Cline, Continue, and five other IDEs since they all speak MCP stdio. You'd reach for this when your agent sessions burn through tokens re-analyzing unchanged context or when you want the agent to learn from mistakes your team already fixed and reverted.