A Field Guide to Memory – Human and Artificial

A field guide to memory — human & artificial

Its context window is working memory — a fixed buffer where thinking happens. When the conversation ends, it’s gone. In human terms that’s severe anterograde amnesia: intact intelligence, no ability to form a lasting memory. Solving it means borrowing from the one system that already did — the brain. This is a guide to how memory actually works, how AI is trying to copy it, and where the ground is still open.

An engram is the physical trace a memory leaves in the brain — a term coined by Richard Semon in 1904Semon 1904, when it was pure theory. A century later, Susumu Tonegawa’s lab found the cells: they tagged the neurons active during an experience, then reactivated them with light and watched the memory returnTonegawa 2012–2015. Engrams are real, locatable, and switchable. This guide is organized around the mechanisms those traces are built from — one specimen at a time.

Each specimen reads the same four ways: how the brain does it, the model that formalizes it, how AI implements or fails it, and the map of who has claimed it. Where the map runs thin, there’s open ground.

Not everything that happens is worth keeping. Something has to decide.

Most of a day is never stored. Novelty, goals and emotional stakes decide what the brain encodes, and nearly every AI memory system tries to make the same call with an importance score.

The moments that mattered are the ones you keep. Feeling is the encoder's thumb on the scale.

the lunch6 monthsstarting strength

fades to half in5 weeks

small talk ▸ starts weak, fades fast · six months later: barely a trace

Arousing moments start stronger and fade more slowly. AI memory systems usually keep only the first half of that.

Forgetting is not the failure of memory. It is one of its functions.

⚖ Contested — Whether forgetting is true decay or retrieval interference is a century-old, still-unsettled question.

spaced repetition: off

a book you read once — what’s left

that perfect line you underlined100%

the argument of chapter one100%

the author's name100%

how it made you feel100%

the gist — it changed how you think100%

drag across the curve

day 0 · curve at 100% · everything is still fresh

Memories weaken with time unless they're renewed: steeply at first, then slowly. It is the most-copied idea in AI memory, and most copies decay in a straight line the brain doesn't follow.

New memories crowd out old ones. Old ones distort new ones. They compete.

Old memories get in the way of new ones, and new ones degrade the old. For AI agents it is the outdated fact that keeps coming back: well measured, barely solved.

Every time you remember something, you change it — and usually make it stronger.

Recalling a memory is not a passive read: it makes the memory stronger and easier to find next time. Every AI memory retrieves; few let retrieval change anything.

Recalling one thing pulls its neighbors with it — everything you learned in the same hour.

Memories formed close together get linked, so recalling one brings back the rest of the afternoon. AI systems copy it with session tags and memory graphs.

Remembering to do something later — the memory that fires itself at the right moment.

A memory for the future: an intention that has to surface on its own when the time or the cue arrives. For AI agents it is mostly open ground.

Memories aren't saved when they're made. They're rebuilt, offline, while you sleep.

⚖ Contested — How consolidation works long-term is actively debated — standard systems consolidation vs. multiple-trace theory.

A memory isn't finished when it forms. The brain replays it, especially during sleep, and slowly folds it into long-term knowledge; AI memory systems copy it with an offline merge pass.

Recall makes a memory briefly editable. Then it's re-stored, sometimes rewritten.

⚖ Contested — The boundary conditions for reconsolidation — when a recalled trace becomes labile — are still debated.

We caught a twelve-inch trout at the lake, and a few people clapped.

Recalling a settled memory can make it editable again before it is re-stored. AI systems do the tame version: update a fact in place and keep the old wording.

"I met her on Tuesday" slowly becomes "I know her" — the event fades, the meaning stays.

⚖ Contested — Whether gist is compressed episodic detail or an independent multi-trace abstraction is unresolved.

Events lose their detail and leave their meaning. AI systems copy the split; distilling gist across many episodes, rather than one, is far less common.

New facts don't land on blank ground. They're folded into what you already believe.

New facts are folded into what you already believe, which speeds learning and bends recall. On the AI side it is close to untouched.

Every patent, paper, and open-source system found across five rounds of adversarial prior-art search. Sorted by priority date, the shape is the finding: the field is occupied, not crowded — a convergence in 2023–2026, with Microsoft, the ex-Cerego team behind Alai Vault, DeWix, Snap, Meta, Samsung, and a fast-moving research frontier all arriving at once.

Foundational predictive modeling of learning effectiveness and memory retention — the root of the Cerego lineage.

The temporal-contiguity effect and the drifting-context account underlying session-based association.

Decades of cognitive-architecture work: prospective memory, affect-modulated declarative memory, base-level activation decay.

Prioritized experience replay weights transitions by salience — reinforcement learning, not LLM memory, but foundational salience-weighting art.

Personalized learning with learner-adaptive scheduling engines.

Personalized learning with automated content generation. Family also includes US11081018B2.

The most on-point Cerego patent: predictive retention/forgetting model with retention thresholds and review scheduling.

Automatically generating related concepts from learning material. Family also includes US11238085B2.

Knowledge-graph memory explicitly framed as modeling the human brain memory system.

Manages focused conversation context, overlaying context retrieved from stored episodes.

Traverses a memory graph; distinguishes reactive and proactive memory for assistant systems.

Generates proactive reminders — a prospective-memory analog for assistants.

Applies relevance decay via linear, exponential, or step functions.

Automatic capture of a searchable store of human memory — the user-facing half of a dual-system memory design.

Later Cerego personalized-learning patent extending the retention-modeling lineage.

Foundational retrieval-augmented generation: feature-vector similarity retrieval feeding an augmented prompt.

Preserves salient entities and facts when summarizing under high compression — pre-compression preservation.

Strongest single §103 obviousness reference. Specification teaches parameterized age-dependent decay of salience, power decay of episodic memories, episodic/semantic tiers, vector-DB long-term memory, a conversational memory tool — though the 20 claims cover brand-marketing agent personalization only. Inventors are ex-Cerego. Family: WO2024178435A1, CA3284082A1.

First-person self-reflective episodic buffer — anticipates first-person memory encoding.

Retrieval-time scoring by recency + importance + relevance — §103 art for dynamic strength and context-dependent retrieval.

Constructs LLM prompts dynamically from a running summary plus a memory-scoring model.

Applies the Ebbinghaus forgetting curve to LLM memory.

Selects context for prompt construction by semantic relevance.

End-to-end chatbot long-term memory: summarize, store, semantic-search, augment the prompt.

Extract-then-cluster consolidation (DBSCAN/HDBSCAN) into memory banks — anticipates two-pass consolidation.

OS-inspired tiered memory (main context vs. archival) with paging between tiers.

Addresses catastrophic forgetting with an episodic/semantic split — for neural-network training, not a retrievable memory store.

Hierarchical topic management that draws session boundaries before the context window fills — granted pre-compression preservation art.

Layered working / episodic / collective memory with embedding-compressed history.

Consolidates a semantic-memory vector DB via DBSCAN clustering with age/effectiveness-based selection — anticipates two-pass consolidation.

Encoding-time memory-conflict resolution: a BERT model detects conflict and resolves by elimination or integration — overlaps part of the interference mechanism (a §103 reference, not anticipation).

KV-cache compression driven by tensor-level frequency + recency importance scoring.

Bounded edge deduplication for complexity reduction in a temporal knowledge graph.

A top-k similarity neighbor set drives memory evolution (Zettelkasten-style linking).

The most directly relevant reference — claims (not just spec) recite the machinery: exponential decay R=e^(-a·x), top-20% summarization, timestamped storage, delete-low-retention (claim 1); importance from parts of speech with adjectives weighted for emotional context (claim 2); ML-derived decay rate (claim 4); semantic-similarity clustering and pruning (claim 5). Claim analysis via machine translation.

ADD/UPDATE/DELETE/NOOP operations with a bounded top-s dedup context. The UPDATE-with-history operation is exactly the in-place-update ('reconsolidation') mechanism — §102 anticipation for it.

A benchmark diagnosing proactive interference as an LLM problem — establishes the term/problem pre-priority, but discloses no mechanism to solve it.

A multi-dimensional numerical salience profile updated by a non-LLM arithmetic rule — §103 art for a salience vector with an affective axis.

The closest research reference. Does contradiction-triggered, encoding-time, graded arithmetic attenuation of memory strength in an LLM agent — most of what a proactive-interference mechanism does. But conflict is LLM-classified (GPT-4o-mini), it attenuates a scalar not a multi-dimensional vector, strength is a stored field batch-rewritten (not computed on read), and it uses one model tier. Predates engram's priority by ~5 weeks.

Retrieval through temporal co-occurrence — §103 art for temporal association.

Conventional CRUD memory store (SQLite + FTS).

Field context only — post-dates engram's priority, so not prior art. Combines proactive interference with sleep-style consolidation.

Field context only — post-priority. Multi-dimensional importance scoring with dual-phase consolidation.

Field context only — post-priority. Hebbian distillation of episodic memories into semantic ones.

Field context only — post-priority OSS memory system.

Interpretation over transcription: the AI writes what it learned in its own words, memories fade unless used, beliefs about people and things are revised only under accumulated pressure and keep their history, and an identity the AI reads at the start of every session. Local: one SQLite file.

If the obvious mechanisms are occupied, where is the real work? Read as a map, the search points fairly clearly.

Almost everyone is building memory systems; almost no one agrees how to measure them. Most results lean on synthetic benchmarks that don’t capture real interaction — the gap behind the RECALL conversation-grounded benchmark.

LLMs demonstrably can’t suppress outdated information. FadeMem and Counterparts both attempt mechanisms — but “the model retrieves a value it should have forgotten” is far better characterized than it is solved.

Nearly every system operates on days or weeks. What a memory should do over months and years — what genuinely consolidates, what becomes irrecoverable versus merely hard to find — is close to unexplored.

Human memory distinguishes “the trace is gone” from “the trace exists but the path is lost.” Most AI systems only delete. Graded, partially reversib