L0
Register
CURRENT ATTENTIONAL CONTENTS
The item is already in the active state. No retrieval. Working memory's last-touched chunk, the CPU's current register file. The answer is already in the answer.
~10⁻¹⁶ J/bit
no decay
The Cascade
Generative recall is the escape hatch.
Most retrieval is solved. Eight tiers handle it deterministically, efficiently, and verifiably. Each tier passes only what it cannot answer to the next. Generative reconstruction — the most expensive option by twenty orders of magnitude — sits at the bottom, not the top. The cascade is what every working biological memory system already does.
The eight tiers
A query enters at L0. Each tier is checked in order. If it answers, the cascade stops and returns. If not, the query falls to the next tier. The total energy cost is the cumulative work of every tier that ran.
L1
Working buffer
RECENT-ACCESS CACHE
A short queue of recently accessed traces. Twenty-second half-life without refresh. Cowan's four-chunk focus of attention, the CPU's L1 cache. Indexed by recency.
~10⁻¹⁵ J/bit
τ ≈ 20 s
L2
Symbolic lookup
INDEX → CONTENT
The query is a key. A B-tree, a hash table, a phonebook, a named variable. Exact match, deterministic, sub-microsecond. The retrieval move computer science solved in the 1970s.
~10⁻¹³ J/bit
O(log n)
L3
Content-addressable
CONTENT → LOCATION
The query is partial content. Hopfield attractor dynamics, Kanerva sparse distributed memory, hyperdimensional similarity. Returns the nearest stored prototype. Hardware CAMs solve this; ML systems should use them too.
~10⁻¹² J/bit
O(1) similarity
L4
Associative recall
CUE → BOUND CONTENT
The query is a cue; the trace is recovered by overlap with what was co-encoded. Semantic memory, paired-associate recall, knowledge-graph traversal. Returns content bound to context, not just content.
~10⁻¹¹ J/bit
multi-hop
L5
Context-reinstatement
RECONSTRUCT ENCODING STATE
The trace can be recovered only by reproducing the surrounding state at the time of encoding. Tulving-Thomson encoding specificity. Hippocampal episodic recall. Every access is a partial replay of the context, not the bit.
~10⁻¹⁰ J/bit
reconsolidates
L6
Consolidation walk
HIPPOCAMPUS → CORTEX
The episode is replayed offline and gradually transferred to the slower, more durable substrate. McClelland-McNaughton-O'Reilly complementary learning systems. Sleep-driven. Returns the slower-but-stable form. The substrate the trace will live on in a year.
~10⁻⁸ J/bit
offline
L7
Generative reconstruction
SYNTHESIZE FROM SCHEMA
The trace is not stored. The query is answered by running a model that produces plausible content from a schema and the query itself. Bartlett 1932. Modern LLM inference. The most expensive retrieval method by twenty orders of magnitude. Should be the last resort, not the first.
~10⁻¹ J/bit
always rewrites
The cost gap
L0 through L4 are six to nine orders of magnitude lower-cost than L7 per bit. The biological brain answers most queries at L0–L5 and reserves L7 for genuine reconstruction problems. Every current AI memory architecture, from KV cache to RAG, runs every query through an L7 process. The cascade is the move available.
Continue