Reference · 15 min read · June 2026
Plain-language definitions of every term that matters in AI memory. Each entry is short enough to quote, long enough to be useful. Designed as the reference you can send anyone who needs to catch up on the category.
The maximum amount of text a model can hold in working memory in one chat.
Measured in tokens. GPT-4-class models: 128k; Claude: 200k; Gemini 2.5 Pro: 1M+. When you exceed it, the oldest turns silently drop.
The unit of text models read and produce — roughly 0.75 of a word in English.
1,000 tokens ≈ 750 English words. Pricing, context windows, and rate limits are all measured in tokens.
Memory that survives between separate chats, not just within one.
Distinct from context window (which resets every new chat). Implemented differently in every tool.
Short notes a model writes about you that get surfaced in future chats.
ChatGPT Memory is the canonical example. Capped, opt-in, opaque about overwrites.
Pattern where the model retrieves relevant documents at query time and includes them in context.
Used to give models access to information that doesn't fit in the context window, like a full knowledge base.
Database that stores embeddings (numerical representations of text) for fast similarity search.
Powers most RAG systems. Pinecone, Weaviate, pgvector are common examples.
A numerical representation of a piece of text capturing its meaning.
Lets you compute similarity between texts — the foundation of semantic search and most RAG pipelines.
Memory of specific events — what was said, when, by whom.
Contrast with semantic memory (general facts). Conversation logs are episodic; a saved fact like 'I'm a designer' is semantic.
Memory of general facts, divorced from the specific moment they were learned.
What ChatGPT Memory tries to extract from your conversations. Easier to store, lossy by design.
Gradual change in how an AI behaves toward you over time, often after model updates.
Especially noticeable in companion apps (Replika, Character.ai). Caused by filter changes, model swaps, or accumulated memory inconsistencies.
When a provider replaces the default underlying model with a new version.
Examples: GPT-4 → GPT-4o → GPT-5. Almost always changes recall and personality even if saved memory is preserved.
When a model version is officially retired and no longer available.
Custom GPTs and prompts tuned to the deprecated model don't behave the same on the replacement.
A reusable ChatGPT configuration with its own system prompt and uploaded files.
Persists across chats but only inside ChatGPT. Not portable to Claude or Gemini.
A scoped Claude container with pinned documents and instructions.
Persistent within Claude. Doesn't auto-learn from conversation; explicit and curated.
Memory entries that a user explicitly marks to keep, exempt from automatic drop-off.
Used by Character.ai, Replika, and others to give users a small set of guaranteed-persistent facts.
The maximum number of saved-memory entries a system will hold.
ChatGPT: a few hundred items. When full, older entries are silently overwritten.
When incorrect or malicious information enters a model's saved memory and influences future responses.
Can happen accidentally (the model misreads your statement) or via prompt injection. Audit memory regularly.
A single distilled unit of personal context — a fact, preference, or pattern — extracted from your raw artifacts.
Konshus-specific term. Atoms are the building blocks of a portable persona.
A raw piece of personal content — a chat export, document, journal entry — before distillation.
Source material for atoms. Lives unprocessed in the vault.
The process of extracting structured atoms from raw artifacts using a language model.
Konshus uses Gemini Flash Lite for distillation to keep cost low; synthesis to a persona uses stronger models.
A portable file representing who you are, derived from your atoms.
Konshus offers four levels: Whisper, Briefing, Full Mirror, Crisis Handoff. Used to onboard any new AI.
Open standard for connecting AI clients to external context sources.
Reached critical mass in 2026 — supported by Claude, ChatGPT custom GPTs, Cursor, and others. First real cross-provider context standard.
Instructions sent to a model before any user message that shape its behavior.
Hidden from the user in consumer apps. Custom GPTs and Claude Projects expose it for editing.
A ChatGPT mode that excludes the conversation from memory, history, and training.
Indicated by a lightning-bolt icon. Easy to start accidentally.
A bulk download of everything a provider has stored about you.
Required under GDPR. OpenAI, Anthropic, and Google all offer one; quality varies.
Permanent removal of data with an audit trail confirming it's gone.
Contrast with soft delete (hidden from your view but retained). Privacy-serious tools should offer hard delete.
Stored data is scrambled such that a storage breach alone can't read it.
Doesn't mean the provider can't read it — they hold the key. Check training-use and access policies separately.
Setting that prevents your conversations from being used to improve the underlying model.
Often off by default on consumer tiers. Check Settings → Data Controls on every provider you use.
Process of further training a model on additional data, including (sometimes) your own.
Distinct from prompting. Most consumer AI memory is achieved via prompting + retrieval, not fine-tuning.
Manually pasting large amounts of background context into every new chat.
Cheap, portable, manual. The original 'personal context document' technique.
When a model is operating on outdated information about you.
Especially common in ChatGPT Memory, where old entries hang around until overwritten.
Collection of a user's writing samples used to mimic their voice in generation.
Konshus stores a voice corpus as part of the persona for tools that imitate your style.
Medium-detail persona export covering your context, preferences, and current focus.
One of four export levels. Goldilocks for handing off to a new assistant without dumping everything.
Minimal persona export — the smallest set of facts that meaningfully shape a response.
Default export. Designed to fit in the system prompt of any model.
Comprehensive persona export including stated identity, patterns, and voice samples.
Counsel+ tier. Designed for high-stakes continuity across providers.
Specialized export for handing off context to another human in an emergency.
Counsel Pro tier. Designed for designated emergency contacts.
Tracking which raw artifact a stored fact came from.
Critical for memory you can trust. Without it, you can't audit or correct entries.
Numeric estimate of how reliable a stored fact is.
Lets a memory system surface high-confidence facts first and flag uncertain ones for review.
Periodic review of stored memory to remove outdated or incorrect entries.
Should be a regular habit — monthly minimum for active users.
The ability for the same personalization to follow you across different AI providers.
Effectively impossible without a dedicated layer; native providers have no incentive to support it.
Konshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.
Meet KonshusKonshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.
Meet Konshus