AI coding assistants are increasingly useful across long-running projects, but their working memory remains tied to individual sessions, products and context windows. Engrim, a new open-source Python tool presented on Hacker News, is designed to separate that project memory from the model or coding environment using it.
The premise is simple: store the durable parts of an AI-assisted development process—architecture decisions, constraints, feedback, current state and next steps—in a local, project-scoped SQLite database. A developer can then switch among supported environments, including Claude Code, Cursor, Windsurf and Google Antigravity, without asking the next assistant to reconstruct the project from scratch.
A memory layer instead of a bigger prompt
Engrim’s core mechanism is a curated memory pack rather than wholesale transcript reuse. At the start of a session, it retrieves a budget-capped context pack; its default documented budget is 4,000 characters. Developers can add records such as decisions, facts, feedback, state and references through a command-line interface or via Model Context Protocol (MCP) tools.
For retrieval, the project combines SQLite’s FTS5 full-text search with local `model2vec` embeddings, then ranks results through hybrid retrieval. It can also run in lexical-only mode, according to its documentation. The design aims to address a practical tradeoff for teams: large context windows may accommodate more history, but repeatedly supplying unfiltered history can increase token use and make important instructions harder for an agent to prioritize.
The database lives locally at `~/.engrim/memory.db`, with owner-only permissions specified by the project. Engrim says it sends no telemetry and uses no cloud synchronization. That positioning will matter to organizations where source code, internal decisions and work logs cannot be placed in a vendor-managed memory system.
Hooks, MCP and an operational workflow
The product is not simply a local notes database. It includes setup commands intended to wire lifecycle hooks into compatible AI coding tools, plus an MCP server for clients that support the protocol. The available MCP actions include recalling memories, adding a record, loading a session context pack and reviewing transcript logs for decisions that may not yet have been captured.
Its proposed operating model is deliberate. Teams or individual developers capture significant choices as they work, write a “resume pointer” before stopping, run a review step to identify uncaptured decisions, and then clear the agent context when appropriate. At the next boot, the tool restores a selected set of records.
Engrim also records an `origin_agent` field, marking whether an item was added through Claude Code, Cursor, Antigravity, the CLI or directly by a user. That provenance is a small but useful feature in multi-agent workflows: it gives a reviewer a clue about where an asserted decision entered the shared memory layer, without treating every agent output as equivalent to a human-approved source of truth.
Why operators should pay attention
The broader opportunity is interoperability. AI developer tools are evolving quickly, and organizations may want to change models for cost, performance, policy or availability reasons. A portable local memory layer could reduce switching friction and limit dependence on a particular assistant’s proprietary history.
It could also encourage a healthier distinction between raw interaction logs and durable project knowledge. Engrim retains a “flight recorder” of turns and actions, but promotes curated records for day-to-day context. That is closer to maintaining lightweight engineering decision records than relying on chat history as documentation.
The important caveat is that the project’s strongest performance claims are self-reported. Its README says it was tested across 105 sessions on a 50,000-line algorithmic trading system and consolidated more than 153,000 tokens of work into a sub-1,000-token memory pack. Those results are promising, but prospective adopters should validate retrieval quality, capture discipline and hook behavior in their own repositories.
What to watch next
For Engrim, the key questions are adoption and reliability: how well it handles conflicting or outdated memories, whether its review process consistently finds important uncaptured decisions, and how smoothly it keeps pace with changing integrations across coding-agent platforms. The project is MIT licensed and distributed through PyPI, making it relatively easy for developers to test.
For builders, it illustrates an emerging architectural pattern: treat AI model context as temporary compute, while keeping the organization’s durable project memory in an independently controlled system.
