What is AI memory?
AI memory is the controlled process of preserving useful information across model calls and retrieving the right parts for a new task. It gives an AI agent continuity beyond one prompt while keeping identity, permissions, relevance, retention, and deletion explicit.
The four layers of useful agent memory
Reliable memory is more than storing text. Each layer limits what can enter a model’s context and why.
Working context
The bounded instructions, retrieved facts, and current task data placed in a model request.
Durable memory
Preferences, project facts, decisions, skills, and lessons that can remain useful across sessions.
Retrieval and packing
Policy-aware selection that turns eligible memories into a concise context pack for one task.
Lifecycle control
Retention, audit, reflection, updates, and tombstoned deletion that keep memory governable over time.
AI memory vs chat history, vector databases, and RAG
| Approach | Best at | Boundary |
|---|---|---|
| Chat history | Conversation continuity | Usually session-shaped; grows quickly and mixes relevant with irrelevant turns. |
| Vector database | Semantic similarity search | Stores and finds vectors, but application policy, identity, lifecycle, and prompt packing remain external. |
| RAG pipeline | Retrieving source knowledge | Grounds a response in documents; it does not by itself model durable agent experience or deletion policy. |
| Agent memory platform | Persistent, scoped continuity | Combines storage, identity, retrieval, context packing, permissions, deletion, audit, and learning workflows. |
Neutron AI
Memory with explicit boundaries
A Neutron Nucleus is an isolated memory universe. Scopes divide it by workspace, user, project, agent, session, policy, or another application boundary. Memory Cells hold minimised facts and lessons; recall and Context Capsules return only eligible, relevant material within a token budget.
Semantic retrieval can supply dynamic knowledge through RAG. Stable context can be compiled and reused through cached context workflows. Deletion creates tombstones so queues, archives, compaction, and cache refresh cannot silently restore forgotten memory.
Consequence workflows add structured scenarios, approvals, observations, and reviewed lessons. They support decisions with uncertainty; they do not expose hidden chain-of-thought or guarantee an outcome.
Where persistent AI memory helps
- Coding agents that remember repository architecture, team standards, past incidents, and verified test commands.
- Customer agents that retain consented preferences and support history without exposing another customer’s data.
- Operational agents that compare new actions with prior decisions, observed outcomes, and bounded consequences.
- Industry agents that retrieve governed domain knowledge alongside organisation-specific policy and evidence.
- Regional or jurisdictional workloads that separate memory by Nucleus, Scope, placement profile, and retention rule.
Start with one bounded memory workflow
Create a Nucleus, choose stable Scopes, remember minimised information, request a bounded agent context pack before the model call, then reflect or forget after the outcome is known.