← @bgSponge

I reverse-engineered Instinct's memory. Here's exactly how it works

Summary

Dhravya Shah shows that Instinct's memory system operates via git-tracked markdown files and grep-style keyword matching instead of vector indexing.

  1. Markdown files power Instinct's memory.

    The core memory architecture relies on git-tracked markdown files, supplemented by specific engineering to ensure speed and seamlessness.

    the memories are stored as git-tracked markdown files

  2. Keyword matching replaces vector indexing.

    Instinct avoids BM25 or vector search, using grep-style queries and file aliases to retrieve information from the file system.

    Instinct quite literally just uses keyword matching / grep-style queries

  3. The agent receives structured context.

    At conversation start, the model is injected with a user profile, a memory one-pager, a compaction recap, and a task board.

  4. Background processes handle memory compaction.

    The agent treats memory as read-only, leaving the work of combining and updating files to a separate background process.

    A background process does the work of combining things

  5. Ingestion follows a daily cycle.

    Memory updates likely run via a 24-hour cron job, as evidenced by a preference taking 23 hours and 16 minutes to commit.

    it's likely a cron job running every day

Highlights

  • Forgetting / decay: Partial: Automatic forgetting missing. Pruning present

  • There's no vector indexing, or BM25 search. Instinct quite literally just uses keyword matching / grep-style queries

  • Instinct is powered by git-tracked markdown files. At it's core, the memories are stored as git-tracked markdown files

  • Supermemory also automatically handles multi-modal ingestion! Below would the full Instinct memory system that works, in supermemory ~exactly like instinct

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →