In-depth analysis · About 5 min read
Sakana’s My Fish makes chat memory inspectable, without proving it correct
The October 15 product update exposes memory facts, source conversations and usage history. Its graph shows when information was used, rather than why an answer is true.
- Published here
- Source published
- Source checked
Sakana AI’s October 15 announcement adds My Fish to Sakana Chat. The feature makes the memory introduced in September visible through an overview, a connection map, a history and links from the conversation itself.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- The Japanese announcement says memory is enabled by default and can be turned off in settings; it selects information useful across conversations rather than saving every detail as a memory.
- In the map, position shows recency, size shows frequency, edges show joint use in an answer, and nodes unused for at least 30 days fade.
What memory the interface is showing
My Fish is an inspection layer for information carried between conversations. Sakana gives preferences, active projects, goals and useful personal details as examples. The Japanese text says the feature does not store every conversation verbatim as memory. That distinction concerns selected memories; it does not describe the retention policy for the underlying chat transcripts.
The same text says memory is on by default and can be switched off in settings. For users, this is a reason to check their current configuration before relying on cross-chat personalization. It is not evidence that turning it off deletes existing facts or source conversations. The announcement presents future improvements in how memories evolve as a direction, not an already measured accuracy result.
From a remembered fact back to its source
The overview groups memories into categories such as profile, preferences, projects and interests. For each memory, the interface shows its fact count, how often it has been read and when it was last used. Opening it exposes individual facts and the conversations from which they were learned, with a route back to the original chat.
The history adds a time-ordered record of reads and writes; a write can be expanded to inspect the facts added. AISubHub’s analysis is that this separates two useful questions: what information is stored, and when did it enter or influence a conversation? It provides an audit path for the user. It does not establish that a fact was interpreted correctly or that every personalization problem is visible.
How to read the connection map
The graph encodes several different things. Recent use moves a memory closer to the center. Frequent use makes a node larger. A node becomes faded after 30 days without use. A line connects memories that appeared together in the same reply. Selecting a node opens its facts, learning time, related conversations and often-associated memories.
These visual rules should not be converted into stronger claims. A large node means frequent use, not importance or truth. An edge indicates co-use, not causation or a model’s internal attention weight. Fading is an age-of-use signal, not a statement that data has been deleted. Those distinctions matter if the graph is used to diagnose why a stale preference keeps affecting answers.
Reviewing memory without leaving the task
During a response, Sakana Chat can show a chip when it reads or writes memory. Selecting that chip opens the corresponding My Fish entry. On mobile, choosing a graph node opens a bottom drawer and shifts the graph to keep the selected memory visible; users can swipe for more details or select another node. The feature is therefore more than a standalone dashboard.
Consider an explicitly hypothetical case: a user once discussed a trip that has since ended, and later receives travel-oriented suggestions during a work conversation. The useful review sequence is to inspect the chip, identify the remembered fact, check its source chat and compare the use history. That is a proposed way to investigate a mismatch, not a scenario Sakana reports testing or a claim that the interface fixes it automatically.
What to verify before treating it as a privacy control
The announced access path is the account menu at the bottom of the sidebar, then Settings, Memory and the My Fish view, on web or mobile. Users can inspect the described records and the memory switch without assuming that visual density or activity implies quality. We did not sign in or inspect anyone’s personal conversations to write this analysis.
Our conclusion is limited: a visible record can reduce the effort needed to investigate personalization, provided the displayed facts and source links are sufficiently complete. A stronger conclusion would need tests of missed memories, incorrect extraction and corrections, plus explicit retention and deletion documentation. This post supplies neither comparative performance evidence nor a new subscription offer. Visibility is a useful step; the remaining behavior must be checked separately.
Why it matters
AISubHub analysis: the value is a traceable route from a personalized answer to remembered facts and their origin. This makes review more practical but does not establish factual accuracy, deletion guarantees or better benchmark performance.
Limits and availability
We read the complete official Japanese and English announcement, not a live private account. The post does not establish memory retention periods, export rights, global eligibility or a price change; fading and disabling must not be described as deletion.
Sources and evidence
Source published: · Source checked:
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.