Why AI forgets important things: JAM and memory for a specific task

A familiar situation: you've been refining a project for a week, you open a new chat — and the AI suggests a solution you've already rejected. A long summary of the previous chat might not help either: a crucial detail might have been lost even while compiling it.

Two papers on a different aspect of memory preparation

In a JAM preprint from September 28, 2026, researchers describe two components. The Memorizer stores full histories in a structured repository with short navigational descriptions. The Researcher, upon receiving a specific question, searches for and reviews the necessary records to gather the answer's context. Thus, the final selection of information occurs once it is already clear why it is needed. JAM: authors' article →

A similar idea, but a different implementation, is proposed by JitMem in a paper from September 23: a specialized component prepares context from selected previous trajectories for the current task. The authors report improvements across three test environments. These research results do not establish an effect for your personal workspace chat. JitMem: separate research paper →

JAM and JitMem are different systems. This is also not an announcement of a new memory button in Claude or ChatGPT. We are analyzing ideas from preprints; the editorial team has not conducted its own comparative launch of these systems.

Example: three records about one form

Imagine a fictional project. On Monday, the team decided to make the phone number mandatory. On Wednesday, the client asked to keep only the email. On Friday, the developer noted that the form was updated, but the mobile scenario had not yet been tested.

To the question "which fields should be left?", the Wednesday decision is required. To the question "can we run ads?", the untested Friday scenario is also important. One general summary saying "worked on the form" provides none of these answers.

Our practical takeaway for small projects: store decisions in a way that allows you to return to their rationale. This is a simplified editorial workflow for managing files, not a reproduction of JAM or JitMem.

What you can organize right now

Create a short description of the current state and a separate decision log with dates. Each entry should answer three questions: what was decided, why, and which previous version it replaces. Do not rewrite yesterday's promise as completed work.

Store actual test results separately: version, scenario, result, and unresolved issues. For the next task, provide relevant fragments or access to specific files. If the assistant does not have access to the archive, simply mentioning its name will not make the information available.

Prompt for checking project memory

Work only with files that I have provided or explicitly granted access to. Current question: [specific question about the project]. First, find records directly related to the question. For each important conclusion, specify the file, date, and the supporting fragment. If records contradict each other, show both and check if there is a later explicit decision. Do not present an old plan as the current state. Do not fill in missing facts with guesses. State what is missing. After answering, compose a short context only for the next task. Do not copy secrets or third-party private data into the general summary.

A small test instead of relying on "long memory"

Take five questions whose answers you know for certain. Among them, there should be a canceled decision, a current limitation, an incomplete check, and a fact that is not in the files at all. The last question tests whether the assistant is capable of admitting a lack of data.

First, provide only a short summary. Then—the same summary along with the required output records. Compare accuracy, grounding, and time spent. This is a local test on your own data, not a scientific benchmark; it will help identify specific gaps in project management.

Access to history also requires boundaries: others' private messages and secrets should not enter the team's shared memory. To start, use our context package template, and before connecting sources, read about AI agent access. The key check is simple: can the assistant show where it got the answer from.

More from this collection

• Claude and formal verification: what was proven for 123 libraries
• Project Swap: AI can bargain, but does it understand what you want?