Associative
Memory Service
Memory is not search. It is a living structure of connected experiences that grows over time. AMS gives AI characters the ability to remember, associate, and recall the way humans do, building genuine relationships through shared history.
What makes it different
- Recalls by association, not by keyword
- Starts with bootstrap memory, not empty
- Grows with every interaction
- Observable, correctable, and safe by design
The Problem
Retrieval is not memory.
Standard approaches find documents that match a query. They do not remember, they do not form associations, and they do not grow. AMS is built on a fundamentally different idea: memory is not a search problem, it is a living structure of connected experience.
| Aspect | AMS | Standard Approach |
|---|---|---|
| Memory Model | Associative graph + semantic + episodic layers | Single text index or context window |
| Recall Method | Graph traversal combined with semantic similarity | Top-K keyword matching only |
| Starting Point | Full life history loaded as bootstrap memory | Empty context, populated from documents |
| Growth | Memory grows and connects with every interaction | Static document store, no evolution |
| Control | User controls their runtime memory: search, edit, delete | No visibility into what is stored or why |
| Privacy | User data separated from character IP, never used for training | User data mixed with training pipelines, no separation |
| Identity | Per-character boundaries enforced in memory | Generic content filter, no character awareness |
| Behavior | Controlled scenario transitions, authored story arcs | Reactive only, no behavioral development path |
| Efficiency | Bounded context, agent gets what is relevant | Full transcript stuffed into context window |
How It Works
Memory in four layers.
AMS combines multiple ways of remembering into a single system. Each layer serves a distinct purpose, and together they create an experience that feels genuinely human.
Episodic Memory
Every interaction is recorded as a living episode. Conversations, events, and observations are stored in chronological order, forming the raw experience stream of the character.
Consolidated Knowledge
Facts and observations are extracted from episodes and consolidated into compact units. The character does not just remember what happened, it understands what it learned.
Associative Graph
Knowledge is structured as a web of connected concepts. When a character recalls a person, it also recalls related places, emotions, and events. Memory navigates by association, not by search.
Semantic Recall
A vector-based retrieval system enables fuzzy, meaning-level recall. Characters can find relevant memories even when the exact words do not match, just like human memory works.
Safety Architecture
Three independent protective circuits.
Characters are adaptive, but never uncontrolled. Three independent safety circuits operate in parallel, each protecting a different aspect of the character and user experience.
Identity Protection
Guards the character's personality and knowledge boundaries. Detects attempts to pull the character out of role, push it beyond its defined knowledge, or make it behave inconsistently with its identity. Responses are checked and softened or blocked before they reach the user.
Scenario Guard
Protects the character's behavioral development path. Authored scenario structures define how a character's story, emotional state, and relationship stages should evolve. The character cannot rewrite, skip, or delete these structures. Every behavioral transition is validated against the authored path.
Memory Scope Lock
Prevents private memory from leaking between users or sessions. Access is computed from authenticated identity and session tokens, never from request body parameters. Your runtime memory cannot mutate the character's configuration, and no other user's memory can reach yours.
Why three circuits: Each circuit is independent and cannot be bypassed by compromising another. A character that stays in role (Circuit 1) and follows its story (Circuit 2) still cannot access another user's memory (Circuit 3). Defense in depth, by design.
Memory Architecture
Three layers. Clear boundaries.
AMS maintains strict separation between three categories of memory. This is not just a technical detail, it is the foundation of how we protect user privacy and comply with data protection regulations.
User Runtime Memory
Memory formed during your interactions with a character: messages, episodes, preferences, facts, vector representations, and graph connections associated with your account. Used for personalization. Fully under your control: view, edit, delete, or export at any time.
Character Configuration Memory
The character's identity, personality parameters, behavioral rules, scenario constraints, and domain knowledge. Authored and editorial content for stable character behavior. Does not contain personal data of any specific user.
Aggregated and Synthetic Experience
General experience structures built from simulation environments, synthetic scenarios, authored content, and reliably de-identified aggregated patterns. No personal conversations or identifiable user data enter this layer. If data cannot be reliably de-identified, it stays in user runtime memory.
Key principle: User runtime memory is never mixed with character configuration or aggregated experience. Deleting your runtime memory removes your personal data without affecting the character's identity or general knowledge.
Capabilities
A memory system, not a storage layer.
Associative Recall
Characters do not retrieve memory by keyword. They walk associative paths between connected concepts, finding relevant memories the way human thought naturally flows.
Bootstrap Memory
Characters start with a full life of memories, not an empty slate. Identity, relationships, experiences, and behavioral history are loaded before the first interaction.
Growing Memory
Every conversation adds to the character's experience. New associations form, existing ones strengthen, and the character's knowledge evolves with each interaction.
User-Controlled Runtime Memory
When you interact with a character, a separate runtime memory layer is formed just for you. This is your personal data: what you discussed, your preferences, shared history. You have full control to browse, search, edit, or delete it at any time.
Character Experience Layer
Characters develop general knowledge through simulation environments, authored content, and reliably de-identified aggregated patterns. This is service configuration and shared experience, not personal data. It cannot be traced back to any individual user.
Behavioral Boundaries
Every character operates within safety boundaries enforced at runtime. Identity protection, topic restrictions, and appropriate response handling are built into the memory system itself, not added as an afterthought.
Scenario-Aware Behavior
Characters do not just react, they develop. Authored scenario structures guide behavioral evolution through controlled transitions. A character's story arc, emotional progression, and relationship stages unfold within defined boundaries that the character cannot override.
Bounded Context
Agents receive a carefully bounded context bundle, not an infinite message archive. The system determines what is relevant for the current moment and delivers exactly that. This keeps interactions fast, focused, and cost-efficient.
Scope-First Memory Isolation
Memory is physically segmented into global, per-user, and per-session scopes. A character's general knowledge, your personal interaction history, and a single conversation session are strictly separated. No scope can leak into another.
Granular Memory Control
Search, inspect, edit, and delete individual memory entities, associations, and entire structures at any granularity. Surgical control over exactly what is remembered and what is forgotten.
Privacy by Architecture
User runtime memory is strictly separated from character memory. Deleting your personal data does not affect the character's identity. User data is never used for model training, never sold, never shared with third parties.
Multi-Brand Isolation
Each product or brand gets its own memory space. Characters in one product never share or leak memory to another, with isolation enforced at the foundation.
Give your characters a mind that grows.
AMS powers every CortexBit character and is available as a memory engine for your own AI agents, robots, or interactive systems.