Your AI forgets you every session. This engine remembers.

What it does

iai-pme is a local MCP server that records every turn of every conversation verbatim and injects a compact, relevant “memory pack” at the start of each new session. It aims to stop your coding assistant from re-asking orientation questions or re-reading files it already saw yesterday. The storage, embeddings, and a Rust-native engine all stay on your machine, with data encrypted at rest using AES-256-GCM.

The interesting bit

The project describes its own memory style as “autistic by design”: exact quotes over paraphrase, precise cues, and rare events kept rare. When facts change, old versions are archived rather than overwritten — the README claims a Rescue@10 score of 1.000, meaning superseded facts remain perfectly retrievable. It also isn’t a wrapper around an existing vector store; the storage engine, community-detection clustering, and hyperdimensional substrate are custom-built.

Key highlights

• Fully local-first: no telemetry, no accounts, no external vector database to install.
• Verbatim, write-once capture: every turn is stored exactly as written, never summarized or rewritten by an LLM during ingestion.
• Token economics: the author claims an injected memory pack costs roughly 88% less in tokens than the agent searches it displaces, with a live dashboard to track your own savings.
• Broad compatibility: speaks MCP-over-stdio with Claude Code, Cursor, Codex CLI, Gemini CLI, Cline, Continue.dev, Zed, and others.
• Native Rust engine with Python bindings; prebuilt wheels ship for macOS Apple Silicon and modern Linux.

Caveats

• Windows support is explicitly beta and the test suite is still being ported.
• Intel macOS has no prebuilt wheel, so the Rust engine must compile from source.
• The ~88% token-savings figure comes from the author’s own three-week usage sample, not an independent third-party benchmark.

Verdict

Try it if you live in Claude Code or Cursor and are tired of repeating yourself every session. Skip it if you need multi-tenant memory for a product you’re shipping — the README itself points to Mem0 or Graphiti for that.