When AI agents move from demonstration to real business, people quickly realize that the inference ability of the model itself is often not the bottleneck, the lack of memory is. An agent who cannot remember the last interaction, accumulate experience, or maintain goals during long tasks finds it difficult to undertake work that requires multiple rounds of collaboration.
The current mainstream agent memory can be roughly divided into three layers: short-term memory, long-term memory, and working memory.
Short term memory usually refers to the contextual window of the current conversation. It determines how much information the agent can 'see' in a task. Due to the limitations of context length and inference cost, simply stuffing all history into it is not cost-effective - it is both expensive and prone to diluting key information. Therefore, a common practice in practice is to slide the window and add a summary: retaining the most recent rounds of original text and compressing earlier content into a summary.
Long term memory spans across sessions and is typically achieved through vector databases or structured storage. Make a trade-off when writing: not all interactions are worth remembering. The common strategy is to first score the relevance and importance, and then decide whether to store it, in order to avoid the memory bank being overwhelmed by noise. When retrieving, it relies on semantic similarity to recall fragments related to the current task and inject them into the context.
Working memory is the "sticky note book" used by agents during a single task execution process. It saves the current sub goals, completed steps, to-do list, and intermediate results. Common paradigms such as ReAct and Plan and Execute implicitly rely on working memory: the model needs to know 'which step I have taken'. When the task chain becomes longer, the management of working memory (compression, deduplication, state rewriting) often has a greater impact on the final success rate than model selection.
A practical suggestion is to first clarify the real needs of the business for memory. If the tasks are mostly one-time short Q&A, the complex three-layer memory architecture will only bring additional operational burden; But if it is a continuous collaboration scenario across days and sessions, then the writing strategy, retrieval quality, and forgetting mechanism of memory should be treated as core engineering issues.
For most teams, starting with "abstract short-term memory+simple vector retrieval" and gradually filling it up based on real failure cases is usually more controllable than designing a large memory framework from the beginning. After all, the value of memory lies not in how much it is stored, but in being able to recall it when it should be remembered.
Comprehensively organize industry information that has been publicly released.