Key Takeaways

• EverMe is evermind.ai's personal AI product. Its Memory Hub uses EverOS, the Apache 2.0 open-source memory runtime, to give connected agents persistent, structured context across sessions.

• The four-layer PBEF model — Profile, Behaviors, Episodic, Foresights — stores categorically different information types, enabling precise retrieval rather than searching one undifferentiated memory block.

• The Profile layer holds stable, low-write semantic facts like occupation and values, while the Behaviors layer continuously updates learned habit patterns and recurring preferences from actual user actions.

• The Episodic layer timestamps specific past events with context. Repeated episodes can inform behavior summaries, but EverMe's public materials do not define a mandatory promotion pipeline between all four categories.

• The Foresights category holds forward-looking suggestions inferred from existing context. A connected agent harness, policy layer, and tools are still required before any action is taken.

• EverOS reports retrieval under 500ms at the p95 percentile as a general memory-engine metric; the public materials do not publish a separate simultaneous PBEF cross-layer benchmark.

• Unlike a plain LLM context window or flat vector RAG, EverMe's layered structure distinguishes stable identity facts from one-time events and learned patterns, preventing outdated preferences from surfacing with the same confidence as current ones.

What Is EverMe and the 'Build a You, Send It to Work' Premise

EverMe is evermind.ai's personal AI product. Its inspectable Memory Hub organizes user context for connected agents, while EverOS provides the underlying open-source memory runtime. That model lets the agent act on that user's behalf without requiring constant re-explanation of context, preferences, or goals.

Most AI assistants reset between sessions. That reset forces users to re-establish who they are, what they care about, and what they are trying to accomplish every time they open a conversation. EverMe eliminates that reset by maintaining a living representation of the user across 4 distinct memory layers: Profile, Behaviors, Episodic, and Foresights.

The premise evermind.ai calls "build a you, send it to work" follows directly from that architecture. The structured context reduces repeated explanations and can help a connected agent reason and prioritize. Execution still depends on the agent harness, permissions, policies, and tools configured around the memory layer.

EverOS is released under the Apache 2.0 license and can be integrated into other agent stacks. EverMe is the product experience built around that memory foundation; the two names should not be treated as interchangeable. The memory system is the foundational layer that separates a personal AI agent from a stateless chatbot: one remembers you, the other does not. For a broader overview, AI agent memory explains how persistent context supports agents across sessions and tasks.

The Four-Layer Memory Model at a Glance

EverMe organizes an agent's memory into four distinct layers — Profile, Behaviors, Episodic, and Foresights — together abbreviated as PBEF. Each layer stores a categorically different type of information. That separation is what allows the agent to reason about a person with precision rather than searching one undifferentiated block of text. There are four layers:

• Profile — stable identity facts: name, role, preferences, relationships, and standing constraints the agent treats as ground truth.

• Behaviors — recurring patterns extracted from repeated actions, revealing how a person characteristically operates across time.

• Episodic — time-stamped records of specific past interactions and events, equivalent to what cognitive science calls episodic memory.

• Foresights — forward-looking inferences derived from the other three layers, representing goals, anticipated needs, and planned commitments.

A memory store without metadata, schemas, or separate indexes may rank facts mainly by recency or similarity. In that setup, there is little structural signal showing whether an item is a durable trait, a habit, a past event, or a future intention. PBEF makes those categories explicit, giving the retrieval layer a clearer way to select context for each task. A memory graph provides another way to organize typed relationships across identities, events, and behavior summaries.

Profile Layer: Stable Identity and Semantic Facts About You

The Profile layer stores the stable, semantic facts that define who a user is: name, occupation, location, native language, dietary restrictions, long-held values, and durable role-based preferences. These preferences do not shift from session to session.

These facts map directly to what cognitive scientists call semantic memory: the brain's store of general, context-independent knowledge about the world and about oneself (source). Semantic memory holds that Paris is the capital of France with the same permanence that the Profile layer holds that a user is a vegetarian software engineer based in Berlin. Neither fact is tied to a specific moment in time; both remain true across contexts.

Profile entries fall into 4 categories of stable attributes. They are: biographical facts (age, nationality, profession), persistent preferences (communication style, dietary needs, language), role-based contexts (parent, team lead, patient), and declared values (privacy-first, sustainability-oriented). Each category supplies the agent with a fixed coordinate it can use without re-deriving it from conversation history.

Profile entries are intended for durable facts and should change when the underlying fact changes — for example, after a career move, relocation, or new dietary requirement. Keeping transient observations out of Profile helps preserve a stable self-model while other categories hold more time-sensitive context.

Behaviors Layer: Learned Patterns, Preferences, and Habits

The Behaviors layer holds dynamic behavioral signatures — patterns, habitual actions, and evolving preferences that EverMe distills from what a user actually does, repeatedly, across sessions. The Profile layer, by contrast, holds static identity facts.

This layer maps directly to what memory science calls procedural memory: the system responsible for encoding habitual routines and skill-based actions that operate below conscious deliberation (source). Procedural memory governs how a person automatically reaches for the same morning workflow or defaults to a preferred communication style without re-evaluating the choice each time. This layer replicates that function for an AI agent.

Three useful categories of recurring signal are habitual action sequences, preference patterns, and avoidance signals. These summaries can inform an agent's decisions, but they do not by themselves authorize the agent to act or guarantee that it will apply a preference in every situation.

Behavior summaries may be refreshed as new evidence appears. The public EverMe materials describe Behaviors as an evolving category, but they do not specify a fixed write frequency or an automatic update rule for every interaction.

Episodic Layer: Time- and Context-Stamped Events

The Episodic layer records specific events and interactions the agent has experienced, each entry stamped with the time and context in which it occurred. The Profile layer holds static facts, and the recurring-patterns layer holds habitual regularities. The Episodic layer holds singular moments — a conversation on a Tuesday afternoon, a decision made under deadline pressure, a preference expressed during a specific project.

This design maps directly to how human memory science defines episodic memory. It is the system that encodes personally experienced events bound to their spatiotemporal context, distinct from semantic memory's context-free general knowledge (source). Endel Tulving, who formalized this distinction, described episodic memory as "mental time travel" — the capacity to re-experience a past event rather than merely recall a fact about it (source).

The Episodic layer applies that principle to an AI agent. Each stored entry preserves the when and where of an interaction, not just its content. This means the agent retrieves not only what was said or decided, but the surrounding circumstances that gave it meaning. A preference noted during a high-stakes client call carries different weight than the same preference expressed casually, and the timestamp-plus-context stamp makes that distinction retrievable.

Episodic entries can provide evidence for later behavior summaries. The public EverMe materials do not specify that repeated events automatically graduate into Behaviors or then propagate into Foresights through a mandatory pipeline.

Foresights Layer: Predictive, Anticipatory Memory for Future Action

The Foresights layer is EverMe's category for forward-looking context, such as anticipated needs, goals, and planned commitments derived from existing memory.

Where the Episodic layer records what happened, Foresights can represent what may matter next. A recurring weekly sync or monthly invoice deadline, for example, can support a forward-looking suggestion when the surrounding application performs that reasoning.

This design is analogous to prospective memory in cognitive science: the capacity to form an intention now and remember it at a future moment (source). The analogy explains the forward-looking category; EverMe's public materials do not define a mandatory trigger-and-action schema for every Foresight.

Foresights can support proactive behavior by supplying forward-looking context to a connected agent. Whether that agent surfaces a reminder, drafts a follow-up, or takes another action depends on its harness, permissions, policies, and tools.

How the Four Layers Interact During Recall and Reasoning

An application can draw from one or more PBEF categories during recall. Each category contributes a different kind of context, and the public materials do not require every query to retrieve from all four simultaneously.

EverOS reports retrieval latency under 500ms at the p95 percentile for its memory engine. That published figure is a general retrieval metric, not a separate benchmark for simultaneous PBEF cross-layer recall. The Profile layer supplies stable identity facts — time zone, role, standing preferences. The Behaviors layer supplies pattern-derived rules — how the user typically responds to a given type of request. An AI memory evaluation framework helps teams compare recall quality, latency, and grounding with consistent criteria.

Episodic records can supply relevant prior events, while Foresights can supply forward-looking context. The application decides which categories to retrieve and how to combine them before generating a response.

EverMe's public product page presents Profile, Behaviors, Episodic, and Foresights as visible memory categories. A practical implementation can use episodes as evidence when updating behavior summaries or profile facts, but the public documentation does not specify a fixed four-step promotion pipeline. Foresights should therefore be treated as inspectable, forward-looking suggestions rather than guaranteed actions.

This distinction matters operationally. Memory can supply better context without retraining model weights, while the surrounding agent remains responsible for deciding whether a suggestion is safe, timely, and authorized to execute.

Grounding EverMe's Layers in Memory Science

Four concepts from memory science offer useful analogies for EverMe's categories: Profile ≈ semantic memory, Behaviors ≈ procedural memory, Episodic ≈ episodic memory, and Foresights ≈ prospective memory (source). These are explanatory analogies, not claims that the software reproduces human cognition.

Cognitive science distinguishes these types because each serves a different retrieval function in the brain (source). Grounding EverMe's architecture in those distinctions means each layer uses a storage and retrieval strategy matched to the nature of its data. Stable facts retrieve differently from time-stamped events. Anticipated intentions, meanwhile, require forward-looking indexing rather than backward-looking recall.

EverMe diverges from the biological model in 2 deliberate ways. First, the layers are explicit and inspectable, whereas human memory types blend unconsciously. Second, Foresights are actively computed from Episodic and action patterns rather than formed through spontaneous intention. This makes prospective memory in EverMe a derived, reasoned output rather than a passive mental note.

How EverMe Differs From a Plain LLM Context Window or Vector RAG

A structured four-layer memory — Profile, Behaviors, Episodic, and Foresights — is what makes EverMe different from a plain LLM context window or single vector store. That structure is what makes a persistent personal agent possible.

A model context window holds only the information supplied to that request. Conversation history persists across sessions only when the surrounding application stores and resubmits it (source). Larger contexts help, but relevant information can still be harder to use when it appears deep inside long inputs (source). A detailed comparison of context window vs agent memory shows why stored context and prompt capacity solve different problems.

A basic vector-store RAG setup often retrieves chunks primarily by semantic similarity. More advanced RAG systems can add metadata, schemas, filters, and separate indexes to distinguish identity facts from events or behavior summaries. EverMe makes those categories visible by default instead of requiring the application team to design that structure independently. Without such structure, an application may surface an abandoned preference with the same confidence as a current one.

EverMe's layered architecture solves 3 distinct failure modes of those approaches:

• Volatility — Profile and conduct patterns persist across sessions without re-injection into every prompt.

• Undifferentiated retrieval — explicit categories let an application apply different filters or retrieval logic to patterns and timestamped events.

• Hallucinated continuity — typed, inspectable memory gives the reasoning engine more structured evidence than an undifferentiated text history; EverOS reports 90.04% recall on HaluMem and 93.05% accuracy on LoCoMo. The LoCoMo benchmark provides a dedicated view of EverOS's reported long-term conversational memory results.

The agent carries identity, habits, history, and anticipation as separate, addressable stores. Not a single undifferentiated pile of tokens. That separation enables autonomous action that remains coherent across days, weeks, and months.

What Layered Memory Means in Practice for Your Personal Agent

A personal AI agent built on layered memory personalizes its outputs, acts without constant instruction, and remains coherent across every session. Profile supplies the stable semantic facts — values, roles, preferences — that anchor every decision the agent takes. Behaviors layer converts those facts into executable habits, so the agent drafts, filters, and prioritizes in the style you have already demonstrated rather than a generic default.

Episodic memory carries the timestamped record of what actually happened. It prevents the agent from repeating a question you answered last Tuesday or missing context that only makes sense in sequence. Foresights translate accumulated history into forward-looking suggestions. A connected agent can surface a task or prepare an output only when its harness, permissions, policies, and tools allow it.

Continuity across time is the practical payoff — the same reset problem that EverMe's architecture is built to eliminate. An agent without layered memory resets at each session boundary, forcing you to re-establish context manually. An agent built on all four layers enters each new session already knowing who you are, how you work, what occurred, and what is likely coming next.

The result is better-grounded delegation. Structured memory helps an authorized agent make more personalized decisions, while execution remains constrained by the permissions and tools supplied to that agent. That is the operational meaning of "build a you, send it to work" — not a chatbot that responds, but a persistent, identity-grounded agent that acts.

Where to go from here

The four layers — Profile, Behaviors, Episodic, and Foresights — form a complete memory architecture. Stable identity facts anchor the agent, learned patterns shape its judgment, time-stamped events supply retrievable history, and predictive foresights drive anticipatory action. EverMind connects this structured memory foundation to the broader EverMe product experience.

Developers who need a portable memory foundation can explore EverOS, evermind.ai's open-source, local-first memory runtime licensed under Apache 2.0. EverOS uses readable Markdown with local SQLite and LanceDB indexes. EverMe presents its own Profile, Behaviors, Episodic, and Foresights organization on top of that foundation; the EverOS README does not describe PBEF as a universal four-stage pipeline. The architecture draws on 5+ peer-reviewed papers covering memory consolidation and retrieval-augmented reasoning, and it is designed to run at production latency. Start building at Build with EverOS Cloud.

Frequently Asked Questions

What do the four EverMe layers each store?

Profile stores stable identity facts — name, role, goals, and long-term preferences. Behaviors stores learned patterns and recurring habits extracted from repeated interactions. Episodic stores time- and context-stamped events, each retrievable by when and where it occurred. Foresights stores predictive entries — anticipated needs and pre-staged actions derived from the other three layers.

How is EverMe's Episodic layer related to episodic memory in the brain?

The Episodic layer directly mirrors the neuroscience concept of episodic memory, which encodes personal events bound to a specific time and context (source). Each entry carries a timestamp and situational metadata. This lets the agent retrieve "what happened, when" rather than abstract facts.

Which memory-science type does each EverMe layer correspond to?

Profile is analogous to semantic memory, while Behaviors is closer to procedural memory because it represents recurring patterns and learned routines (source). Episodic corresponds to episodic memory — event-bound, time-indexed records (source). Foresights corresponds to prospective memory — the cognitive function of remembering to act at a future moment (source).

Why does EverMe use four separate layers instead of one large context window?

Four dedicated layers let an application retrieve the category a reasoning step needs instead of placing an entire history in every prompt. EverOS reports <500ms p95 retrieval for its memory engine, but the public materials do not attribute that result to a specific PBEF routing mechanism or publish a universal token-cost reduction for every workload.

How is EverMe's structured memory different from a vector-store RAG setup?

A basic vector-store RAG system may retrieve chunks mainly by embedding similarity. More advanced RAG systems can add metadata and typed indexes. EverMe provides Profile, Behaviors, Episodic, and Foresights as explicit categories, so an application can route or filter retrieval by memory type instead of designing that taxonomy from scratch.

What is the Foresights layer and how does it let an agent act ahead of your needs?

The Foresights layer stores forward-looking context informed by existing memory. A connected agent may use a Foresight to prepare or execute an action, but the reasoning workflow and execution controls must be configured separately through its harness, permissions, policies, and tools.

How does a single interaction move through the EverMe layers over time?

An interaction can be stored as an Episodic record. Later consolidation may use repeated episodes to refine a behavior summary or profile fact, while forward-looking suggestions can be stored under Foresights. The public EverMe documentation does not state that every interaction must propagate through all four categories in sequence.