Glonce

Engrim gives AI coding agents a shared local memory across tools

toolsThe Glonce editorsHacker News🔥57Glonce rating57Impact45Novelty60Relevance85Surprise35Players20Credibility80
Engrim gives AI coding agents a shared local memory across tools

Tim Gordon posted engrim on Show HN: a local-first, project-scoped SQLite memory engine meant to let developers move between different AI coding agents, Google Antigravity, Claude Code, Cursor and Windsurf, on the same project without losing architectural decisions, constraints or state each time a session is cleared or a tool is switched. The pitch is that as context windows grow past 1 million tokens, reloading a project's full history on every turn causes what the README calls attention dilution: reasoning degrades and cost rises with each turn, while clearing context causes total amnesia. Instead of resending that history, engrim keeps a curated episodic memory pack capped at 4,000 characters, built by combining SQLite FTS5 keyword search (bm25) with static vector embeddings from model2vec in a hybrid reciprocal-rank fusion engine. Each stored memory carries an origin_agent field, so entries created by Antigravity, Claude Code, Cursor or the CLI stay attributable when several agents work the same repository. Installation is a single pip install engrim, followed by engrim setup, which detects which of Antigravity, Claude Code or Cursor are present and wires each one's hooks, status line or MCP configuration automatically; per-tool setup commands and a --dry-run flag are also available, and Windsurf is configured by hand through its MCP config file. The README reports one production test: across 105 continuous sessions on a 50,000-line algorithmic trading codebase running real capital, the tool ran with zero regressions across 186 unit tests and no context amnesia across model switches, and it consolidated over 153,000 tokens of accumulated architecture, tuning and debugging work into an active memory pack under 1,000 tokens, under 1% of the context window, which the author describes as a 99%-plus cut in reloaded context cost on every session restart. Data stays local in a SQLite file at ~/.engrim/memory.db, with no telemetry or cloud sync, database files created with owner-only 0600 permissions, and *.db paths gitignored by default. The project is MIT-licensed and at version 1.3.0. At the time of writing the Show HN had drawn 26 points and 5 comments.

Key facts

  • engrim is a local-first SQLite memory engine that lets AI coding agents, Google Antigravity, Claude Code, Cursor and Windsurf, share project decisions and state on the same codebase.
  • It keeps a curated episodic memory pack capped at 4,000 characters, retrieved with a hybrid of SQLite FTS5 keyword search and model2vec static vector embeddings.
  • The author's single reported production test, on a 50,000-line algorithmic trading codebase across 105 sessions, cut a 153,000-token work history to under 1,000 tokens, a claimed 99%-plus reduction in reloaded context.
  • Setup is pip install engrim followed by engrim setup, which auto-detects installed agent environments and wires their hooks or MCP configuration; Windsurf is added manually to its MCP config.
  • Memory is stored only locally at ~/.engrim/memory.db with no telemetry or cloud sync, database files use 0600 permissions, and the project is MIT-licensed at version 1.3.0.

Why it matters

Developers who use more than one AI coding agent on the same project currently re-explain architecture and constraints every time they switch tools or clear a session; engrim's pitch is to externalize that state into a local database so any supported agent can pick it up, turning a full context reload into a lookup of a few thousand characters.

Who it affects

Developers who work across multiple AI coding environments on one codebase, specifically those using Google Antigravity, Claude Code, Cursor or Windsurf, and anyone concerned about vendor lock-in from keeping project history inside one tool's proprietary session state.

How to use it

Install with pip install engrim, then run engrim setup to auto-detect and configure whichever of Antigravity, Claude Code or Cursor are present on the machine; Windsurf requires adding engrim by hand to its MCP config file. Commands like engrim add and engrim review capture and check decisions as work happens, and clearing a session with /clear does not lose memory, since engrim reloads the active pack on the next prompt. It is MIT-licensed and requires no account or cloud service.

How solid is it

The only evidence offered is a single self-reported production case, one 50,000-line algorithmic trading codebase run by the author across 105 sessions and 186 unit tests, with the 99%-plus context reduction figure derived from that same case. There is no independent benchmark, no comparison to other memory or vector-database tools for coding agents, and the Show HN post itself had drawn modest traction, 26 points and 5 comments, at the time of writing.

Risks and caveats

The performance and reliability claims rest on one author-run codebase rather than broader or independent testing, so results on other projects or languages are unproven. The project has a single named maintainer, Tim Gordon, with no stated team or company behind it, which matters for long-term maintenance. No pricing or paid tier is mentioned; it is presented as free and local-only, with all data kept in an unencrypted local SQLite file that individual users are responsible for securing.