AI Memory Types: Short-Term, Long-Term, and Shared Records

AI memory types are ways of describing what an assistant retains, how long it remains available, and where it lives. Short-term memory usually concerns the current conversation; long-term memory preserves selected information across sessions. Other labels describe the contents, such as facts or instructions, so these categories can overlap rather than compete.

A shared project record such as fenbs belongs to the external, durable side: tasks, approved rules, and progress stay outside any one chat. The AI memory library covers the larger topic, while this guide helps you interpret the labels.

Which distinctions are most useful when choosing memory?

Ask three separate questions. Scope: who and which conversations can use the information? Content: does it describe facts, past events, or instructions? Storage: is it built into the AI app, held in an external record, or reflected in the underlying model? A product can have several answers to each question.

These labels are useful working descriptions, not a universal certification. A vendor’s “long-term memory” might save a short summary, while another system stores individual records. Ask how each system selects, updates, and retrieves information. The name alone does not establish accuracy, portability, or access control.

What separates short-term context from lasting memory?

The context window is the material a model can use while producing its current answer. Anthropic’s context documentation↗ (opens in a new tab) distinguishes that working material from the model’s training. A conversation can therefore feel coherent now without becoming a reliable record for a different session tomorrow.

Long-term memory means selected information remains available beyond that immediate exchange. It still needs to be found and supplied when relevant. “Saved” and “read for this answer” are different states. A user may be able to open an old conversation even when the assistant’s current answer does not include it.

For example, an assistant can compare several draft headlines during one conversation. The lasting record may only need the approved headline, the reason for choosing it, and where it will be used. Keeping the entire discussion can be useful as history, but it should not blur the difference between suggestions and the final choice.

What do semantic, episodic, and procedural mean?

LangChain’s memory overview↗ (opens in a new tab) uses these three labels for facts, experiences, and ways of doing things. They borrow from human-memory language; they do not establish that an AI remembers like a person. A practical classification describes the record’s purpose without making claims about consciousness or understanding.

• Semantic: a fact or concept, such as which approved design the project uses. Keep its source and scope so a later assistant can check it.
• Episodic: a record of an event, such as a test that failed after a configuration change. Preserve what happened and what was actually observed.
• Procedural: instructions for carrying out work, such as checking message delivery before calling a form ready. Keep the steps distinct from proof that they were followed.

A single lesson may combine all three. It can describe a failed delivery test, the fact learned about a configuration, and the check to perform next time. You do not have to force each entry into one exclusive box. Keep the event, conclusion, and future instruction clear enough to review separately.

In fenbs, Lessons learned and Decisions and rules give these different records a practical home. A lesson records what experience taught the team; a rule records a person’s instruction that continues to apply. One should not silently become the other.

What does native memory tell you about storage?

“AI native memory” can mean memory built into an AI application, but the phrase does not explain the mechanism by itself. It may refer to stored preferences, generated summaries, or a product’s own memory tools. Ask what survives a new chat, what can be edited, and whether another supported app can read it.

An external record has a separate location: a document, database, task tool, or another connected service. That makes a different boundary visible. The assistant must receive or read the relevant record, and access depends on the connection and the person’s permissions. External storage does not automatically imply shared access or perfect recall.

Information reflected in model training is another category again. It is not the same as an editable project entry with an owner and a source. For ordinary project work, use records that can state an approval, show where it came from, and change when the person’s decision changes. The AI brain introduction explains this deliberate record of work.

How do personal and shared records differ?

Personal memory helps one person’s assistant respond appropriately. A project record describes a particular body of work. A shared record makes selected information available to the right colleagues and their assistants. These audiences can overlap, but a useful private preference should not automatically become a rule for everyone.

Suppose a designer prefers concise explanations, while a project requires complete accessibility notes. The preference can guide conversation style; it should not cause the assistant to omit required project evidence. Identify which instruction controls the task and which simply makes the interaction more comfortable.

AI memory for teams covers the responsibilities that come with a shared record. When switching apps is the issue, portable AI memory addresses how the record reaches another assistant. Both questions are separate from whether a memory contains facts or experiences.

How to do this with fenbs

1. Create a free account and add an active project with its current tasks. One person can start free with no credit card.
2. Separate lasting facts in context notes, approved choices in Decisions and rules, and useful experiences in Lessons learned. Include source details rather than copying every conversation.
3. Connect your supported assistant from Integrations. MCP is a standard way for an AI app to read and update an outside tool.
4. Ask the assistant to read the project context at the start and save task progress plus Where we left off at the end. Check a fresh session against the saved next step.

fenbs does not train a model or import your personal chat memories. It fits a deliberate project record; a developer building a memory engine needs other infrastructure. The AI memory graph guide explains a separate storage structure for relationships rather than another mutually exclusive type of memory.