Unified contextual memory · v1.0
The conversation that lasts years.
Conventional LLMs suffer from contextual amnesia: once the window fills up, everything before it vanishes. In Crisol, memory is not a store you query — it is native structure ingested directly in the forward pass. No amnesia. No RAG bolted on the side.
The impossible dilemma of the context window.
Every transformer carries the same structural problem: the context window is finite and its cost grows quadratically.
Small window
8K or 32K tokens: the model forgets everything that came before. Today's assistant has no memory of yesterday's conversation.
Giant window
1M or 10M tokens: prohibitive cost, diffuse attention and more hallucinations. Solutions like RAG help, but they are external hacks around the model, not native memory.
Crisol does not choose between forgetting and paying. It solves the problem by architecture: memory built into the runtime, not hung off the side.
Cohesive architecture
Three layers that compose on every step.
Crisol's memories are not external, queryable stores. They are integrated as three structural layers that assemble every time the organism reasons.
Core — HolographicCore
The compressed cognitive vector of the moment. Goal, state, constraints, entities, decisions and a causal summary in a structure small enough to fit in the CPU's L1 cache. Identical across every tier: the map that keeps the organism oriented at all times.
Cortex — SlotMemoryAllocator
Immediate cognitive RAM. Context fragments with a time-to-live, typed as working, conversation, mission or evidence. It does not replace persistent memory: it is the working cortex that complements the holographic core.
Stores — long-term memory
The stores for all 8 canonical memory types, held in RAM with optional persistence to disk. The frame keeps only references by ID — never heavy contents. This is where the days, months and years of the conversation live.
The 64 KB HolographicCore is identical across every tier. What scales is the cortex: 16 slots in Mini, up to 96 in Maestro.
Eight canonical memory types.
Each type plays a distinct role in the organism's cognitive life, from the volatile slot of the current second to the causal graph that builds itself while the Crisol rests.
Working
Live slots holding context fragments with a TTL. The immediate cortex of the present moment.
Episodic
What happened: events and turns with scope, recoverable when they matter.
Prioritized episodic
The high-utility episodes the organism chooses to retain above the rest.
Semantic
Knowledge crystallized from recurring, consolidated patterns.
Procedural
Sequences and procedures: how things are done, not just what happened.
Conversational
The thread of each session, turn by turn, with no cross-contamination between conversations.
Identity
Organism scope: who this Crisol is, its persistent character beyond any single session.
Causal
The graph from automatic causal dreaming — nodes, edges and verdicts about how the user's world works.
Memory ↔ model cohesion
Not queried. Ingested.
The three layers assemble into a UnifiedMemoryFrame. The difference from RAG is radical: the model does not receive memory as an attached text it has to read. It receives it fused into its own cognitive state during the forward pass.
And all of this costs less than 2% of total compute. Memory is not a toll: it is part of how the organism thinks.
HoloCore → initial state
Infuses the 64 KB HolographicCore into the contextual state of layer 0. The model does not start blank: it starts with the user's living memory already reconstructed.
Sparse attention to live slots
A globally shared memory expert attends, token by token, to the live slots of the frame at every layer. Memory takes part in the routing, not outside of it.
Consolidate what matters, forget what doesn't.
Every fragment travels through an explicit, auditable pipeline. What is useful rises a layer; what is irrelevant is retired. Like a healthy mind, Crisol forgets on purpose.
The fragment enters unfiltered, freshly captured from context.
It proves recurring usefulness and becomes a candidate for consolidation.
It crystallizes into a higher layer: episodic, semantic or procedural.
The forgetting policy retires it through expiration, irrelevance or contradiction.
A controlled forgetting policy
Forgetting is not an accident: it is a policy with three triggers — expiration (unused beyond its time-to-live), irrelevance (usage below threshold over a window) and contradiction (later verified information refutes it). Every decision is recorded and auditable.
The conversation that lasts years
From shift assistant to life companion.
500 turns on a technical project
Each turn is stored as a fragment in a live slot. The HolographicCore captures the key decisions in 64 KB. On close, everything is serialized to disk if persistence is enabled.
Pick up as if it were yesterday
Crisol loads the project stores and rebuilds the 64 KB core. The CoreEncoder infuses that core into the contextual state of layer 0. The conversation continues with no amnesia.
50,000 accumulated turns
The consolidator has crystallized the patterns into semantic memory. The irrelevant was forgotten by policy. Automatic causal dreaming has already built a full causal graph of the project.
Switch projects, lose nothing
The old project's stores are exported as a Custodia memory package. They stay archived and recoverable. The slots are freed for the new work. Nothing is lost by default.
This is not possible in any conventional LLM. It is the difference between an assistant that resets every morning and an organism that remembers alongside you.
Remembering is your decision, not the system's.
Persistence to disk is optional by design. Anyone who values maximum privacy can run 100% in volatile RAM: when it powers down, no trace remains. Anyone seeking continuity enables persistence and keeps years of conversation.
The system does not impose — it proposes. Sovereignty over one's own memory always belongs to the Crisol's owner. Everything is further isolated by hierarchical scopes: each project and each session live apart, with no cross-contamination between conversations.
The next step
Living memory feeds the agents that reason.
The core, the slots and the stores do not exist to archive: they exist so the agentic runtime can decide, plan and verify with the user's full context present at every step.
Explore the agentic runtime →