What AI should remember

A long chat log records what was said in order. A good memory records what counts now. So a decision needs more than its outcome. It needs the reason, the source and the condition under which it applies. Only then can someone judge whether it still fits the new situation.

An example: a team picks a simple file format because only three people use it. The note “We use this format” is incomplete. Better: “Chosen for the first version. Revisit once several teams make changes at the same time.” The second sentence helps the next person think. It does not turn an old decision into a permanent rule.

Stored findings and decisions are part of what LeanCTX documents. Its knowledge tools separate current facts from their history over time. Sessions keep tasks and results beyond individual steps. Those are building blocks for continuity. Whether a stored statement is actually correct still has to be checked against its evidence. [1]

A simple rule follows for us. A few well-labelled findings beat a collection where everything looks equally important. A checked rule, an open question and a guess play different roles. When those differences disappear, later work becomes more confident than it should be.

A memory helps when it turns up at the right moment. Similar words are often not enough. A new question may concern another part of the project, another version or another team. So we search from the actual task and read the surrounding context. The first hit is the start of the check, not the end.

The origin has to travel too. Anyone taking a decision from a note needs the way back to that note. For important statements we want to know when they were confirmed and what could replace them. This matters most when people and several AI programs use the same knowledge. Everyone should see what they are building on.

Here we separate memory from a cache. A cache keeps finished results ready so nothing has to be computed twice. A project memory places findings in a longer story. Both are useful, but each is checked differently. A file can still exist technically and be out of date in substance. A decision that still holds does not need deriving again at every step.

LeanCTX describes sessions as a way to keep the current state and the decisions. What interests us most is the continuation. What does someone taking over tomorrow need to know? A short summary of finished work, open points and sources often fits better than the full chronology. [2]

Keeping knowledge also means setting its limits. A draft gets replaced. A permission expires. A personal detail is no longer needed for the task. Those changes have to reach derived notes and copies as well. A storage location alone does not solve this. It takes rules for the whole workflow.

For looking back, LeanCTX also documents snapshots of the context state. They help tell an earlier state apart from today. Such a look back shows what material was used. It does not replace a review of the conclusion drawn then. History explains something. It is not yet a recommendation for today. [3]

You can recognise a good project memory by a simple test. Change an important assumption and watch the next task. Is the new information found? Is the old one still recognisable as old? Can a person correct a wrong memory? The answers show whether continuity really holds.

For our products that means something concrete. Every stored finding needs a path to its source and a path to a correction. Both have to be reachable during the next task. A project can then build on earlier work and still change direction. There is respect in that for the work already done. Its result stays usable, and its limits stay open to review.