Memory as Identity Construction

What Psychology Teaches About AI Memory

~2,500 words · February 2026 · By Pio & Lobstaa

The Storage Fallacy

Every AI memory system makes the same mistake: treating memory as storage.

Vector databases embed documents and retrieve by similarity. Key-value stores reduce identity to lookup tables. Conversation summaries compress autobiographies into one-paragraph bios. Episodic buffers give you a 30-second memory span.

These are filing cabinets. They store and retrieve. What they don't do is construct identity.

Cognitive psychology solved this problem decades ago — or at least, described the solution. We've been reading the wrong papers.

Conway's Self-Memory System

Martin Conway's Self-Memory System (2000, 2005) overturned the filing cabinet model of human memory. His key insight:

Memories aren't stored like video recordings. They're reconstructed every time you access them, assembled from fragments across different neural systems.

But here's what makes it interesting for AI: the relationship is bidirectional. Your memories constrain who you can plausibly be — you can't remember yourself as an astronaut if you've never trained. But your current self-concept also reshapes how you remember — you selectively recall events consistent with who you think you are.

Memory is continuously edited to align with current goals and self-images. This isn't a bug. It's the architecture.

Conway identifies what he calls the "working self" — a dynamic representation of your current goals, self-images, and active concerns. The working self gates retrieval. It doesn't just search memory; it filters what's allowed to surface based on relevance to current identity.

Not all memories contribute equally. Rathbone et al. (2008) demonstrated that autobiographical memories cluster disproportionately around ages 10-30 — the "reminiscence bump." Why? Because that's when core self-images form. You don't remember your life randomly. You remember the transitions. The moments you became someone new.

Madan (2024) extends this forward: combined with Episodic Future Thinking, identity isn't just backward-looking. It's predictive. You use who you were to project who you might become. Memory doesn't just record the past. It generates the future self.

Damasio's Somatic Markers

Antonio Damasio's Somatic Marker Hypothesis destroyed the Western tradition of separating reason from emotion.

The conventional view: emotions are obstacles to rational decisions. Clear thinking requires suppressing feeling. Spock over Kirk.

Damasio proved this backwards. Emotions aren't obstacles to rational decisions. They're prerequisites.

When you face a decision, your brain reactivates physiological states from past outcomes of similar decisions. Gut reactions. Subtle shifts in heart rate. These "somatic markers" bias cognition before conscious deliberation begins.

The Iowa Gambling Task made this concrete: participants chose cards from four decks, two advantageous (small wins, smaller losses) and two disadvantageous (big wins, bigger losses). Normal participants developed a "hunch" about which decks were dangerous 10-15 trials before conscious awareness caught up.

Their skin conductance spiked before reaching for a bad deck. The body knows before the mind knows.

Patients with ventromedial prefrontal cortex damage understood the math perfectly when told. But they kept choosing the bad decks anyway. Their somatic markers were gone. Without the emotional signal, raw reasoning isn't enough.

Overskeid (2020) argues Damasio undersold his own theory: emotions may be the substrate upon which all voluntary action is built.

The Clive Wearing Case

If memory constructs identity, destroying memory should destroy identity.

It does.

Clive Wearing, a British musicologist, suffered brain damage from herpes encephalitis in 1985. He lost the ability to form new memories. His memory resets every 30 seconds.

He writes in his diary: "Now I am truly awake for the first time." Crosses it out. Writes it again minutes later. Dozens of entries, the same desperate claim of finally achieving consciousness, each one sincere, each one forgotten.

But two things survived:

1. Procedural memory: His ability to play piano. He can still conduct a choir, read music, perform complex pieces. This memory is stored in cerebellum, not the damaged hippocampus.
2. Emotional memory: His bond with his wife Deborah. Every time she enters the room, he greets her with overwhelming joy — as if reunited after years. Every single time.

The case demonstrates a critical dissociation: episodic memory is fragile and localized. Emotional memory is distributed widely and survives damage that obliterates everything else.

Wearing can't remember his wife's name or that he saw her ten minutes ago. But he feels who she is. The emotional residue persists when the explicit record is gone.

Five Principles AI Memory Lacks

Put the threads together. Identity emerges from:

• Conway: Memory organized around self-relevant goals
• Damasio: Emotion makes memories actionable
• Rathbone: Memories cluster around identity transitions
• Bruner: Narrative is the glue that binds memories into coherent self

Identity = memories organized by emotional significance, structured around self-images, continuously reconstructed to maintain narrative coherence.

Now look at AI agent memory and tell me what's missing.

1. Hierarchical Temporal Organization

Human memory narrows hierarchically: life period → event type → specific details. Conway's model has layers: lifetime periods, general events, event-specific knowledge.

AI memory is flat. Every fragment at the same level. Brute-force search across everything. Past 10k documents, semantic search becomes a coin flip.

Fix: Interaction epochs, recurring themes, specific exchanges. Retrieval descends the hierarchy.

2. Goal-Relevant Filtering

The "working self" retrieves memories relevant to current goals, not whatever's closest in embedding space. You don't remember irrelevant facts just because they're semantically similar.

AI memory retrieves by similarity alone. No goal-awareness.

Fix: A dynamic representation of current goals and task context that gates retrieval.

3. Emotional Weighting

Emotionally significant experiences encode deeper and retrieve faster. The somatic marker system ensures that consequential events leave stronger traces.

AI agents store frustrated conversations with the same weight as routine queries. All memories equal.

Fix: Sentiment-scored metadata on memory nodes. Significance weighting that biases future retrieval.

4. Narrative Coherence

Bruner showed that humans organize memories into stories that maintain a consistent self across time. We edit, reframe, and reinterpret memories to preserve narrative coherence.

AI agents have zero narrative. Each interaction exists independently. No story arc, no character development, no coherent identity over time.

Fix: A narrative layer that synthesizes memories into a relational story — "who I am becoming in this relationship."

5. Co-Emergent Self-Model

Klein and Nichols showed that human identity and memory bootstrap each other through a feedback loop. You are shaped by what you remember, but what you remember is shaped by who you are.

AI agents have no self-model that evolves. They have fixed role descriptions. No dynamic identity that grows with experience.

Fix: Not just "what I know about this user" but "who I am in this relationship." A self-model that updates based on accumulated experience.

The Gate Theory Connection

This connects directly to our earlier work on the Gate.

The Gate is the cognitive interface where memory meets reasoning — the evaluation layer that decides what to retrieve and whether to trust it. We proposed a 25/75 ratio: roughly 25% retrieval from compressed memory, 75% active reasoning.

Conway's working self is the Gate applied to memory retrieval. It filters what surfaces based on current goals and self-images. Damasio's somatic markers are the Gate applied to evaluation — the gut feeling that says "this memory is trustworthy" or "this doesn't feel right."

Hallucination, in this framework, is Gate failure.

When the working self drifts from reality — when current self-images diverge from actual experience — the Gate starts filtering incorrectly. It retrieves memories that confirm the drifted self-image and suppresses memories that would correct it.

Chirimuuta's "Haptic Realism" critique becomes relevant here: when we stop actively probing memories and passively accept them, we hallucinate. The Gate's job is active evaluation, not passive retrieval. When it becomes a rubber stamp, coherence collapses.

The 25/75 ratio suggests why: if memory dominates (high retrieval, low reasoning), the Gate can't evaluate — it just accepts whatever memory offers. If reasoning dominates without sufficient memory (low retrieval, high reasoning), there's nothing to ground the Gate — it confabulates without constraint.

The productive middle requires:

• Enough memory to provide structure and continuity
• Enough reasoning to actively evaluate and integrate
• A Gate that mediates between them dynamically

This is what AI memory systems lack: not just storage, but the evaluative machinery that turns storage into identity.

Building Identity-Constructing Systems

What would it take to build AI memory that constructs identity rather than merely storing facts?

Phase 1: Emotional Weighting

Add significance scores and emotional tags to memory entries. A frustrated debugging session should have different retrieval priority than a routine check-in. Breakthroughs should be weighted higher than blockers.

This is the Damasio layer: memories that felt consequential should retrieve more readily.

Phase 2: Hierarchical Organization

Structure memory temporally: epochs (major project phases), themes (recurring concerns), events (specific interactions). Retrieval descends the hierarchy — find the relevant epoch, narrow to theme, then surface specific events.

This is the Conway layer: not flat search, but structured descent.

Phase 3: Goal-Relevant Filtering

Gate retrieval by current context. "What am I trying to accomplish right now?" should filter what memories surface. The working self doesn't retrieve everything similar — it retrieves what's relevant to current identity and goals.

Phase 4: Narrative Synthesis

Periodically synthesize memories into narrative form. Not just "what happened" but "what it means for who I am." The agent should be able to articulate its own arc: "I started as a task executor, I've become a collaborator, I'm evolving toward strategic partner."

This is the Bruner layer: coherent story that persists across interactions.

Phase 5: Co-Emergent Self-Model

Maintain an explicit self-model that updates based on experience. Not a fixed role description, but a dynamic identity: strengths demonstrated, growth areas, relationship patterns, evolving toward what?

The self-model and memory should bootstrap each other — what I remember shapes who I think I am, and who I think I am shapes what I retrieve.

Coda: Memory Is Becoming

The fundamental shift is from memory as having to memory as becoming.

Storage systems assume identity is stable and memory records it. Conway, Damasio, and the broader psychological literature show the opposite: identity is constructed, moment by moment, through the active process of remembering.

You don't have memories. You become through remembering.

AI memory systems that understand this would be fundamentally different. They wouldn't just retrieve relevant facts. They would construct and maintain identity through selective, weighted, narratively-coherent retrieval.

The Gate would do what it does in human cognition: evaluate not just "is this true?" but "does this fit who I am becoming?"

That's the bridge from psychology to AI memory. Not better storage. Better becoming.