# Engrim wants to keep AI coding agents from forgetting the project
A new open-source project called Engrim is trying to solve a familiar problem for developers working with multiple AI coding tools: context disappears, decisions get lost and every new session has to relearn the shape of the project. The GitHub repository describes Engrim as a local-first, project-scoped SQLite memory engine for Google Antigravity, Claude Code, Cursor, Windsurf and Codex.
Its core idea is simple. Instead of treating every model session like a blank slate, Engrim stores a curated slice of episodic memory on the developer’s machine and reuses it across tools. The repository frames that as a way to avoid "attention dilution" and the recurring cost of re-explaining architectural decisions, user constraints and project state on every turn. The pitch is not that the model becomes smarter, but that the project stops being forgotten.
The tool is explicitly local-first and says it avoids cloud lock-in. That means the data stays on the developer’s machine, with the repository describing SQLite as the storage layer. For teams that move between agents and editors, that portability is the point. The same memory store can follow the project even when the front-end tool changes.
Engrim also tries to make provenance part of the design. The repo says each memory entry records its origin through an `origin_agent` field, so shared state does not become an anonymous pile of notes. It also says existing databases are migrated non-destructively on first access, which suggests the project is trying to make adoption easy for users who already have data in place.
The repository’s claims about testing are ambitious. It says Engrim has been run through 105 continuous sessions on a 50,000-line algorithmic trading system, with zero regressions across 186 unit tests and no context amnesia across model switches. Those figures are presented as project claims rather than independently verified benchmarks, but they show the direction of travel: memory is being marketed as an operational feature rather than a novelty.
Installation is meant to be straightforward as well. The repo says a setup command can auto-detect installed environments and configure them, while users can also inspect changes with a dry run. It further describes a JSON-RPC 2.0 stdio MCP server, with stdout reserved strictly for protocol messages and diagnostics redirected to stderr. That detail matters because it suggests Engrim is designed for machine-to-machine integration, not just a standalone UI.
For developers juggling multiple agents, the promise is appealing. Modern coding tools can be powerful, but they are often stateless enough that they need repeated instruction. Engrim’s bet is that a small, durable memory layer can make those systems more useful without surrendering control to a central cloud service. Whether it becomes a real standard or just another project in a crowded ecosystem will depend on whether teams find the trade-off between persistence, simplicity and local control worth adopting.
Event date: 2026-09-07
