Show HN: Engrim – A universal, local-first SQLite memory engine for AI CLIs

Does it come with an uninstall script?

I've been using another memory plugin [1] for the past couple months. Since this change, I almost never run session compaction. Instead, I opt to just make a new session, give the agent a task, and have it figure out what happened in previous session(s).

Decision summaries look really cool -- though the agent/model info probably won't be useful for me because I don't actually ever let agents commit code autonomously (even in branches).

- [1] https://ctx.rs/

I especially like the provenance idea.

> Empirical Proof (The 105-Session Case Study)

Is this something you can benchmark against competitors? Lots of these memory plugins are cropping up and nobody knows how they compare.

It seems the meat of your work is in https://github.com/timgordontg/engrim/blob/main/src/engrim/m...

There is no lifecycle management or conflict resolution yet, and the models do the work of deciding when to use it.

I built engrim to establish a local-first, open standard for cross-model AI agent memory.

As context windows scale past 1M+ tokens, developers face rapid attention dilution: reasoning degrades, and token costs multiply exponentially with every turn. But if you clear your agent's session (/clear) to save money and speed things up, the agent suffers total episodic amnesia, forgetting architectural rules, past debugging steps, and micro-decisions.

Engrim replaces attention dilution with a 4,000-character curated episodic working memory pack. It decouples your project's intelligence from single-vendor proprietary cloud silos. You can switch seamlessly from Gemini in Google Antigravity to Claude 3.7 in Claude Code to GPT-4o in Cursor or Windsurf mid-project—your agents pick up exactly where the others left off.

A few architectural details: - Under the hood, it's a zero-latency hybrid retrieval engine combining SQLite FTS5 (BM25 keyword search) with static vector embeddings (model2vec) using Reciprocal Rank Fusion (RRF). - Memory retrieval is gated per prompt and filtered by a relevance floor, meaning only high-signal records enter your token window. - Provenance Tracking: It maps the origin of every memory entry via an `origin_agent` field (antigravity, claude-code, cursor, cli, user) across multi-agent setups. - 100% Local & Offline: Runs entirely out of a local SQLite database (~/.engrim/memory.db) with POSIX 0600 file permissions and zero cloud telemetry.

Empirical Proof: I production-tested this across 105 continuous sessions on a 50,000-line algorithmic trading system. Over 153,000 tokens of architecture and parameter-tuning logs were consolidated into an active memory pack under 1,000 tokens. That represents a 99%+ cut in reloaded context costs on session restarts with zero architectural regression.

Quickstart: It configures environment lifecycle hooks automatically (e.g., configures hooks.json for Antigravity, settings.json hooks and CLAUDE.md for Claude Code, and registers the stdio MCP server for Cursor and Windsurf):

pip install engrim engrim setup --all

I'm hoping this helps developers escape cloud lock-in and keep their data sovereign while putting an end to massive token bills. I'd love to hear your thoughts on the schema approach, the hybrid RRF engine, or how you handle episodic state across different AI tools!

If multiple agents are reading and writing to the same memory, I'd love to be able to see which agent created a memory, which agents later retrieved it, and how it propagated across sessions. A timeline or knowledge graph of that could be really useful, especially for tracking down stale or conflicting memories.

How do you handle memory eviction when context windows get large? Are you doing semantic similarity cutoffs or just recency? The 80-line constraint is impressive—did you consider supporting structured recalls (e.g., "all conversations about X topic")? Or is that out of scope for the minimalist angle?

Building this locally vs. cloud-hosted changes the whole game for AI CLI tools. Would use this.