# Supermemory reverse-engineers Instinct's memory down to git, Markdown and grep > Source: > Published: 2026-09-21 04:19:28+00:00 # Supermemory reverse-engineers Instinct's memory down to git, Markdown and grep **Dhravya Shah's black-box teardown says the personal agent relies on daily files and keyword search, then rebuilds the pattern in 60 lines.** By [Ryan Merket](https://runtimewire.com/author/ryan-merket) ยท Published Primary source: [X](https://x.com/DhravyaShah/status/2101745550752428340) ## Why it matters Instinct's reported design shows that useful agent memory can rest on simple storage. The defensible work sits in consolidation, retrieval and deciding what an agent should remember. [Dhravya Shah (@DhravyaShah)](https://x.com/DhravyaShah) spent three days probing Instinct and concluded that the personal AI agent's memory runs on a surprisingly conventional stack: git-tracked Markdown files, keyword search and a background process that periodically reorganizes what the agent knows. Shah, the founder of [supermemory](https://supermemory.ai/), published the findings [in a thread on X](https://x.com/DhravyaShah/status/2101745550752428340) on September 20th and provided the fuller technical account in [a company blog post](https://supermemory.ai/blog/reverse-engineering-instinct-memory/). He also wrote an approximately 60-line implementation that recreates the reported pattern using supermemory's API. The simplicity is the point. Instinct appears to get useful long-term recall without placing a vector database or semantic retrieval system at the center of every query. The harder engineering sits around the files: deciding what deserves to be saved, reconciling contradictions, building a compact user profile and feeding the right material back into the agent when it acts. Shah has spent the past three years working on agent memory. He previously worked on AI infrastructure at Cloudflare and started supermemory as an open-source consumer project before turning it into developer infrastructure. In October 2025, supermemory said it raised a $3 million round led by [Susa Ventures](https://www.susaventures.com/), with Browder Capital, SF1 and several angel investors participating. The teardown also arrives days after Shah [discontinued supermemory's company brain and Nova products](https://supermemory.ai/blog/an-update-to-supermemory), effective September 9th, to concentrate on the underlying memory engine. That gives the research a clear commercial purpose: show developers that a sophisticated personal agent can use a legible memory architecture, then offer supermemory as the managed version. ### Files, aliases and a daily cleanup job According to Shah's reconstruction, Instinct organizes information into directories for people, organizations, durable facts, preferences, decisions, communications, timelines and active workstreams. Individual Markdown records contain structured headers, aliases and links to related records, producing what Shah described as a loose graph. The aliases appear to make basic keyword retrieval more dependable. Shah reported that a search for "pasta" or "takeout" found a dining-preference record, while a semantically similar phrase such as "Italian noodles I enjoy" did not. Misspellings also failed. From those tests, he concluded that Instinct uses grep-style matching rather than vector search or BM25 ranking. Instinct also appears to inject a compact profile before answering. Shah measured one snapshot at roughly 4,250 tokens, covering life context, communication style and the circumstances under which the agent should act without asking. A separate conversation recap occupied about 8,750 tokens. Todos and the current session identifier supplied additional working state. Shah inferred that a background ingestion process runs roughly once per day after watching a preference take 23 hours and 16 minutes to appear in a reported commit. That process seemed capable of moving temporary details into workstreams, replacing incorrect facts with dated corrections, shortening records and retaining earlier versions in git history. Instinct also created daily and weekly memory files. Those findings remain a black-box reconstruction. Shah did not inspect Instinct's source code, and Instinct did not confirm the architecture in the materials reviewed. The model responsible for producing the profile, the reconciliation prompts, conflict handling and the precise storage layer remain unknown. ### The harness carries the hard part Shah's tests suggest that the underlying storage format explains only part of Instinct's performance. The agent still needs tools for listing files, searching records, inspecting history and managing tasks. A separate process must decide how conversations become durable records without distracting the model that is handling the user's immediate request. That separation matches Shah's own product thesis. He argues that the primary agent should read memory while background systems handle consolidation and writes, which become increasingly expensive as records accumulate and conflict resolution grows harder. His assessment of Instinct was favorable, with qualifications. Shah found strong single-fact recall, handling of corrections, recency and explicit personalization. Multi-step recall across sessions was weaker, automatic forgetting appeared incomplete and multimodal personalization performed poorly in his tests. He did not test performance over months of use or more than one million tokens. The 60-line example at the end of Shah's post recreates the overall flow with a wrapped language model, automatically injected profiles, memory-search tools, dynamic profile buckets and one conversation document per day. It does not reproduce Instinct's undisclosed ingestion pipeline or profile-generation system. Those components are where much of the operational complexity remains. ### Two founders betting on persistent agents Instinct founder [Noah Shinn (@noahrshinn)](https://x.com/noahrshinn) has worked on the same problem from the agent side. Before Instinct, Shinn was an early researcher at Sierra. His published work includes Reflexion, a framework that lets language agents retain written reflections from previous attempts, and tau-bench, a benchmark for agents interacting with users and software tools. Shinn introduced Instinct publicly in August as a personal agent users can text or call. Instinct is designed to operate phones and computers on a user's behalf, covering tasks such as email, calendars and bookings. That scope makes persistent memory central to the product: an agent handling purchases, appointments and personal communications needs a usable record of preferences, relationships and unfinished work. Shah's reconstruction suggests that Instinct's advantage may come from disciplined orchestration around ordinary storage primitives. Git and Markdown make records inspectable and versioned. Aliases make crude search useful. Background consolidation keeps the active context manageable. The resulting system is simple enough to explain, while the decisions governing what gets written, recalled or forgotten remain difficult to copy.