Context icons

The 53 context icons in AI & LLM, described

Belief statebelief-state

A ring with a head over a line — what the agent currently takes to be true.

Checkpoint statecheckpoint-state

A ring with a peg in it — the state saved so a run can pick up from here.

Compress contextcompress-context

The context window with chevrons pressing inward — the same content in fewer tokens.

Context budget alertcontext-budget-alert

The context window with an exclamation mark — the budget nearly spent.

Context cachecontext-cache

The context window with a block kept in it — tokens the model does not have to read again.

Context compresscontext-compress

Context compression — fitting more meaning into fewer tokens by summarising or squeezing.

Context fullcontext-full

Context full — there is no room left in the model's window for more tokens.

Context pincontext-pin

The context window with a pin in it — a piece of context that stays no matter what is dropped.

Context prefixcontext-prefix

The context window with a fixed first line and a point for the rest — what every request begins with.

Context routercontext-router

One line splitting into two over a tray — each request given the context it needs.

Context windowcontext-window

The context window — how many tokens a model can see and reason over at once.

Conversation bufferconversation-buffer

The context window with three lines in it — every turn so far, kept as it was said.

Entity memoryentity-memory

A tag landing in a tray — what the agent remembers about each named thing.

Event-sourced agentevent-sourcing-agent

A ring with three events in a row — state that is nothing but the events replayed.

Fact storefact-store

A statement with a full stop over a tray — the facts an agent may cite.

Graph memorygraph-memory

Three points joined over a tray — memories kept as things and the links between them.

Hierarchical memoryhierarchical-memory

Rows narrowing upward over a tray — a little kept close, more kept further away.

Memory auditmemory-audit

A checklist over a tray — what the agent remembers, gone over by a person.

Memory conflictmemory-conflict

Two points with a wall between them over a tray — two memories that cannot both be true.

Memory decaymemory-decay

A line fading in steps over a tray — memories that count for less as they age.

Evict memorymemory-evict

An arrow leaving over a tray — a memory pushed out to make room.

Forget memorymemory-forget

A cross over a tray — a memory the agent is told to drop.

Memory indexmemory-index

A hash over a tray — the index that finds a memory by what it is about.

Memory mergememory-merge

Two lines joining into one over a tray — memories about the same thing folded together.

Memory namespacememory-namespace

Brackets over a tray — memories kept apart by which agent or user they belong to.

Promote memorymemory-promote

A chevron lifting a bar over a tray — a memory moved to a tier that is kept longer.

Memory quotamemory-quota

A level between two posts over a tray — how much an agent may remember.

Memory retrievalmemory-retrieve

A lens over a tray — a memory looked up and brought back.

Memory tiermemory-tier

Bars stepping down over a tray — hot, warm and cold memory.

Note to selfnote-to-self

A speech bubble with a pin in it — something the agent tells itself to remember.

Pin factpin-fact

A pin over a tray — a fact that stays in memory whatever else is dropped.

Prefix cacheprefix-cache

The context window with a first line and a bolt under it — the shared prefix, already computed.

Prompt cache hitprompt-cache-hit

The context window with a bolt in it — the prompt found already computed.

Prompt cache missprompt-cache-miss

The context window with a slash through it — the prompt not in the cache, computed from scratch.

Replay logreplay-log

A clipboard with a play mark on it — the log an agent's run can be replayed from.

Replay stepreplay-step

A ring with a skip mark in it — one recorded step played again.

Restore staterestore-state

A ring with an arrow looping back — the state put back to what it was.

Rolling summaryrolling-summary

The rotation loop with lines in it — a summary rewritten as the conversation goes on.

Scratchpad notescratchpad-note

A clipboard with a pencil stroke on it — working notes the agent keeps while it thinks.

Session statesession-state

A ring with two lines in it — what is true for this session and no other.

Shared memoryshared-memory

Two points tied together over a tray — one memory several agents read.

Sliding attentionsliding-attention

The context window with an arrow sliding through it — attention that looks only at the last stretch.

Agent snapshotsnapshot-agent

A ring with an eye in it — a look taken at everything the agent holds, kept as it was.

Stale memorystale-memory

A clock over a tray — a memory older than the thing it describes.

State diffstate-diff

A ring with a plus over a minus — what changed between two states.

State lockstate-lock

The context window with a lock in it — state no other run may change while this one holds it.

Agent state machinestate-machine-agent

A ring with two states and the step between them — an agent that is always in exactly one.

Summarise turnsummarise-turn

The context window with a line boiled down to a point — a turn kept as its gist.

Token budgettoken-budget

The context window with a coin in it — how many tokens a call is allowed to spend.

Truncate contexttruncate-context

The context window with three lines, the last cut short — what falls off the end.

Turn windowturn-window

The context window with two turns in it — how many exchanges the model can still see.

Vector memoryvector-memory

An arrow over a tray — memories stored as vectors and found by nearness.

World modelworld-model

A ring with a triangle in it — the agent's own picture of how things work.

In code, each is one import — import { BeliefState } from "@iconmind/react/icons/belief-state" — and the same name in Vue, Svelte, Solid, Preact, React Native, Astro, Blade and Flutter.

Narrow it by tag