Engrim expands local SQLite memory to four AI coding assistants

Why developers need persistent project memory

Coding agents such as Claude Code and Cursor rely on the model’s context window to retain reasoning.

When a developer clears a session, the conversation history – and the rationale behind generated code – is lost.

Timothy Gordon, a researcher in quantitative finance, AI and machine learning, describes this loss as a common frustration.

He notes that aggressive clearing of sessions often erases the “why” behind design choices.

Engrim 1.3.0 adds a shared, local memory layer

On September 7, Gordon released Engrim 1.3.0, extending his open‑source memory engine to Claude Code, Cursor, Windsurf and Google Antigravity.

The update turns a tool originally built for Claude Code into a cross‑agent memory layer that works through native hooks or a Model Context Protocol (MCP) server.

Engrim stores curated memories in a SQLite database on the developer’s machine, keeping the data under the user’s control.

When a new session starts, Engrim retrieves a small, relevant subset of records and continues to search the store as prompts arrive.

Installation requires only two terminal commands: pip install engrim and engrim setup.

No hosted account, API key or external service is needed, preserving privacy and reducing dependency on third‑party platforms.

The project’s thesis is explicit: “The models are disposable utilities; your project’s decisions are not.”

Developers can record decisions—such as selecting PostgreSQL over MySQL—or reject an authentication scheme and retain the justification across agents.

Each memory entry logs the originating environment, allowing later sessions to reload decisions made by any of the four supported agents.

How Engrim manages memory without overwhelming the model

Engrim distinguishes between curated memories—facts, feedback, preferences, references—and an append‑only transcript log that acts as a flight recorder.

The transcript preserves activity without automatically dumping the entire history back into the model’s context.

This separation prevents the problem of an ever‑growing archive that forces the model to sift through stale or irrelevant material.

Engrim builds a budget‑capped memory pack by combining SQLite FTS5 keyword retrieval with local static embeddings.

It merges the two ranking signals using reciprocal‑rank fusion to surface the most pertinent records.

The MCP server exposes commands for adding records, recalling them, assembling a context pack and reviewing recent transcripts for unsaved decisions.

Users may mark older records as superseded, preserving the audit trail while keeping the active memory concise.

A resume pointer identifies the next task for the following session, supporting continuity across breaks.

Before a session is cleared, the “engrim review” command scans recent activity and warns if a likely decision has not been saved to curated memory.

Gordon’s approach counters the industry trend of expanding context windows, which gives coding agents more room but does not solve the loss of reasoning when sessions are reset.

By keeping project knowledge locally and independently of any model, Engrim offers a durable knowledge base that survives agent switches and model updates.

The open‑source nature of the project invites community contributions and transparency around how memory is managed.

Version 1.2.2, uploaded on August 12, laid the groundwork for Claude Code integration, while 1.3.0 adds the multi‑agent capability.

Engrim’s design demonstrates that durable, user‑owned memory can be layered on top of existing AI coding workflows without requiring proprietary services.

Developers who adopt Engrim can maintain a coherent architectural narrative even as they experiment with different coding assistants.

This capability may influence how future AI‑assisted development tools handle session persistence and knowledge transfer.

As AI labs continue to push context‑window limits, solutions like Engrim illustrate an alternative path focused on local, curated memory.

Why This Matters