Instinct's reported memory relies on keyword matching, not semantic search, with a once-a-day write job

Keyword search is cheap at query time and expensive at write time. Every phrasing a user might later type has to already be in the file, which is what the aliases in each record header are for ^[4]. Shah's probes found the edge. Searches for "pasta" and "takeout" both reached a dining-preference record, the phrase "Italian noodles I enjoy" did not, and misspellings failed as well ^[5]. Nothing normalises the query, so the semantic step happens once, whenever the aliases get written.

Those writes run on their own clock. Shah watched one preference take 23 hours and 16 minutes to appear in a commit and inferred a background ingestion pass running roughly once a day ^[7]. He reported that the same pass moves temporary details into workstreams, replaces incorrect facts with dated corrections, shortens records and leaves earlier versions in git history ^[8]. A correction therefore becomes durable when the pass runs, not when the user says the old fact was wrong.

The recurring bill is in the injection, not the search. Shah measured the profile snapshot at roughly 4,250 tokens, covering life context, communication style and when the agent may act without asking, with a separate conversation recap at about 8,750 ^[6]. That is about 13,000 tokens of context ahead of the todos and the session identifier ^[16]. The cost is fixed per turn whether or not the question touches any of it, and it stays affordable only because the consolidation pass keeps shortening records ^[8].

For the pattern to transfer, several things have to hold. The corpus has to belong to one person, small enough that grep across it is fast and that a summary of it fits in a context window. Users have to tolerate up to a day between saying something and the system being able to retrieve it. And the harness needs tools for listing files, searching records, inspecting history and managing tasks, because the agent does its own retrieval ^[10].

The reconstruction is black-box. Shah did not inspect Instinct's source code, Instinct did not confirm the architecture in the materials reviewed, and the model that produces the profile, the reconciliation prompts, the conflict handling and the exact storage layer are unknown ^[9]. He also has a position in the answer. Supermemory discontinued its company brain and Nova products effective September 9th to concentrate on the memory engine ^[12], raised a $3 million round in October 2025 led by Susa Ventures with Browder Capital, SF1 and angels ^[13], and published a roughly 60-line reimplementation of the reported pattern on its own API ^[3]. Sixty lines is short enough to read end to end, which is more than most architecture claims offer.

Shah's assessment of Instinct was favorable with qualifications, and he reported strong single-fact recall and handling of corrections ^[15]. His own thesis is that the primary agent should read memory while background systems handle consolidation and writes, and that those writes get more expensive as records accumulate and conflict resolution grows harder ^[11]. On the evidence of one user's notes, I would accept the narrow version of that: with good aliases, grep retrieves well enough, and the difficulty moves to the process that maintains the files.