Start free
The vocabulary around persistent memory is young and used loosely. These are the definitions we work from — written to be useful whether or not you use My-CoMind.
The ability to resume any topic exactly where it was left — with the decisions in effect, the questions still open, and the reasoning behind them intact.
The emphasis is on resumption rather than retention. A complete archive that requires an hour of reading before you can contribute again has retained everything and provided no continuity. The test is how quickly someone can act, not how much was stored.
See also: the full explanation, context reconstruction.
The observable sequence of positions held on a single topic over time, with the interval between each shift.
Drift is descriptive. Changing position is usually evidence of learning, so treating it as a defect — or scoring it — would be both wrong and quickly resented. What matters is that a shift stays visible: both positions preserved, with their dates, so the duration of the previous view is legible.
It answers a question no summary can: not what do we think, but how did we come to think it, and how settled is it. A position stable for a year and one re-litigated four times may read identically in a document; their drift histories do not.
A single captured unit of thinking — a meeting, a document, a dictated note — kept whole.
The unit matters. Splitting one conversation into a dozen fragments optimizes for retrieval and destroys the thing that made it coherent: a decision separated from the discussion that produced it loses the reasoning that justifies it. One source becomes one memory.
A view of one topic that stays current as new material arrives, showing where things stand, what was decided, and what remains open.
The contrast is with a folder, which holds documents but has no opinion about them, and with a summary, which is accurate on the day it is written and decays from then on. A workspace is regenerated rather than maintained, which is what allows it to stay true without anyone being responsible for updating it.
Treating a decision as an object with status and lineage, rather than as a sentence inside a document.
A decision recorded as prose cannot tell you whether it is still in force. A decision recorded as an object can: it knows when it was made, what it replaced, and whether something later replaced it. This is the difference between an archive that accumulates contradictions and one that resolves them.
See also: superseded decision.
A decision replaced by a later one, kept with a link to whatever replaced it.
Deleting the earlier decision is the intuitive move and the wrong one. The superseded version carries information the current one does not: that the question was contested, what the alternative was, and why it was abandoned. When circumstances change, that record is what prevents a team from rediscovering an option they already rejected for good reason.
A question raised and not yet resolved, carried forward until it is answered.
Open questions are the most valuable and most fragile artefact in any body of work. They are typically recorded once and then silently stop being tracked — not because they were resolved, but because no mechanism carried them into the next conversation. A question that disappears without an answer is indistinguishable, later, from one that was never asked.
The accumulated context an organization holds independently of whichever individual happens to remember it.
Most organizations believe they have institutional memory when what they actually have is a few long-tenured people. The distinction becomes visible during leave, growth, or departure — the moments when the question "why do we do it this way?" stops having an answer.
The work of rebuilding the state of a topic before being able to contribute to it.
This is the cost cognitive continuity exists to remove. It is rarely measured because it is distributed in small amounts across many people — the fifteen minutes before a call spent skimming old notes, the message asking a colleague where something landed, the meeting that opens with ten minutes of recap for one attendee.
It is worth estimating for your own team. The per-instance cost is small enough to ignore and the aggregate is usually not.
See also: the full article, with cited research and a calculator.
A representation of material as connected entities rather than as isolated documents.
The practical consequence is that relationships become queryable. In a folder structure, the fact that two projects hit the same obstacle is invisible unless someone noticed and wrote it down. In a graph, the shared entity is the connection, and it exists whether or not anyone noticed.
An assistant's ability to retain and reason over your material across sessions, rather than starting from nothing each time.
Frequently conflated with a large context window, which is a different thing. A large window lets a model consider more at once within a single exchange; persistence is about what survives between exchanges, and about the selection problem — deciding which of thousands of memories are relevant to the question actually being asked.
The amount of text a language model can take into account at once.
A capacity limit, not a form of memory. Once the exchange ends, the contents are gone. Growing windows have narrowed the gap enough that the two get confused, but the distinction holds: a window is how much can be read at a time, memory is what is still there tomorrow.