Barcelona software developers using Claude Code, Cursor, Codex CLI or other MCP-compatible coding assistants can now assess iai-personal-memory-engine, an open-source tool intended to preserve working context between sessions. For people repeatedly explaining project decisions, coding constraints or previous work to an assistant, the practical aim is to make that information available at the start of later conversations while keeping the memory store on their own machine.

CodeAbra describes iai-pme as a local server using the Model Context Protocol, or MCP, a standard that lets AI tools exchange context with compatible applications. The project is offered under the MIT licence and lists support for Claude Code, Claude Desktop, Cursor, Codex CLI, Gemini CLI, Cline, Continue.dev, Zed, Cherry Studio, Goose, Aider, Hermes, OpenClaw, Le Chat and Kimi.

“Your AI assistant forgets you every session. iai-pme gives it a memory that doesn't.”

Local storage and compatibility

According to the project's documentation, the engine captures conversation turns verbatim, organises them over time and returns a selected memory pack when a new conversation begins. It says the engine, storage and embeddings run locally, with no account, API key or telemetry required.

That local-first design may be relevant to Barcelona-based freelance developers, small software teams and students who want to test persistent AI-assistant context without sending the memory layer to a separate cloud service. The normal model request made by the user's existing command-line tool still leaves the device, the documentation says.

The project states that macOS and Linux are fully supported. Windows support is described as beta: the runtime has been validated on Windows 11, while the test suite is still being ported.

Token-saving figure is a developer benchmark claim

CodeAbra says the tool's ambient memory packs cost about 88% fewer tokens than the agent searches they replace. Its repository documentation says that figure was measured on the author's own store over three weeks: 282 memory packs injected about 99,000 tokens of context and, under the project's conservative formula, displaced about 707,000 tokens of agent searches.

The project says it supplies a benchmark harness so users can run the tests themselves. It distinguishes those automatically injected packs from an explicit memory_recall request during a session, which has a default budget of 1,500 tokens and is not claimed to achieve the same 88% reduction.

How developers can test it

The repository provides a macOS and Linux bootstrap script that checks prerequisites, installs the background engine and capture hooks, registers an MCP server with Claude Code, and runs a health check. Developers who want to inspect the process before making changes can add --dry-run; --preflight-only checks prerequisites only.

  • Use the project's bootstrap script for the guided macOS or Linux installation.
  • Claude Code users can use the repository's plugin route, then restart their session for automatic capture and recall.
  • Check the project's dashboard after installation to view memory packs served, injected tokens and its estimate of tokens saved.

The repository had 662 GitHub stars and 79 forks when the supplied source material was collected.

Primary sources: github.com, CodeAbra. Reported by raw.githubusercontent.com, relatedrepos.com, dev.to, Scott Dallen.