Dhravya Shah, founder of Supermemory, said his reverse engineering of the personal agent assistant Instinct suggests the product handles long-term memory with Git-managed Markdown files instead of stuffing full chat histories back into context.
According to Shah, Instinct organizes long-lived information into a file-based system and keeps an updated view of the user’s current state. Supermemory itself is built as long-term memory infrastructure for agents. Shah said it extracts facts from chats, files, and web pages, maintains relationships between people, tracks changes over time, and builds user profiles, then passes relevant memory to agents through APIs, MCP, and related interfaces.
What Shah says he found in Instinct
After breaking down Instinct, Shah said the two systems appear to share a very similar idea of how an agent should remember things.
In his black-box testing, Instinct appeared to receive a user state summary of about 4,250 tokens at the start of each conversation. That summary included life background, operating preferences, and communication habits.
More detailed records — including person-specific facts, preferences, decisions, and tasks — seemed to remain inside Markdown files and were pulled in only when needed through keyword search similar to grep.
Shah said he did not find evidence of vector indexing or BM25 retrieval.
Memory appears to be reorganized in the background
Shah said the memory store is also being continuously reworked behind the scenes. Temporary information is grouped into specific tasks. Experiences are condensed into long-term preferences. New facts overwrite or revise older ones. Git is used to preserve version history.
In that setup, the main agent appears to focus mostly on reading, while the actual organization and merging work is handled by backend processes. Shah estimated that this maintenance cycle runs about once per day.
Supermemory could largely reproduce the design, Shah says
Shah said the architecture can already be largely replicated with Supermemory. In the example he shared, roughly 60 lines of code were enough to reconnect user profiles, memory search, historical versions, and continuous updates to an agent.
Earlier, Reflexion co-author Ashwin Gopinath had speculated that Instinct’s edge might come from continuously maintaining a "current state" that can track what is still valid, what has expired, and what remains unfinished.
Based on Shah’s analysis, that state may not rely on any mysterious database. It may instead sit on top of a file system that is being continuously organized.