Atlas Librarian · CMI v2
Adaptive recall asks which memory is useful for this cognitive state—not merely which text looks similar.
The next generation of the Conscious Memory Index introduces an additive adaptive-recall path for Atlas Librarian. It evaluates a wider set of eligible memory candidates, considers their cognitive role and current lifecycle, and assembles a bounded Working Memory for the task. The established fast-recall path remains available while CMI v2 is evaluated and strengthened.
Algorithmic direction
The memory architecture now has a policy layer that can optimize what becomes active.
Segmented memory provides structure. CMI v2 adds a selection policy across that structure. The public description identifies the signal families and observable behaviour, while the exact ranking recipe remains proprietary.
A rule, procedure, handover, decision, event and durable fact are not treated as interchangeable passages. The memory type contributes to its fitness for the current task.
Higher-risk tasks can require a stronger preference for governed or authoritative memory, while lower-risk reasoning can admit more contextual material.
Freshness, thermal state, actual use and supersession help distinguish currently useful knowledge from material that remains valuable only as history.
A due follow-up or a session restart can change which memory classes deserve priority even when the underlying text has not changed.
Links among memory items can contribute context across cognitive segments without collapsing them into a single generic store.
When an operational result is explicitly verified, that evidence can inform future recall as one bounded signal rather than allowing retrieval frequency alone to define usefulness.
Value per unit of context
Persistent intelligence is useful only if active cognition stays selective.
CMI v2 treats Working Memory as a scarce reasoning resource. Candidate memory can be considered without all of it becoming active context, and only the final selected set is counted as checked out by the adaptive path.
The Librarian may inspect a larger eligible candidate pool before deciding what should participate in reasoning.
Optional memory competes for a finite Working Memory budget. Mandatory governed context can be preserved even when it is not the cheapest context to load.
Candidate consideration is separated from final selection so the system does not teach itself that every inspected memory was genuinely used.
A frequently retrieved memory is not automatically a good memory. Verified successful, failed or rolled-back outcomes can provide stronger evidence about future usefulness.
Implementation status · 5 September 2026
CMI v2 exists as an additive internal evaluation path; it is not being presented as a finished universal benchmark winner.
The current CMI v2 alpha implements state-aware candidate scoring, authority and risk sensitivity, lifecycle and supersession handling, relationship signals, verified outcome evidence and bounded Working Memory assembly. Speedway is retaining the existing recall path in parallel while comparative evaluation expands.
Adaptive candidate ranking, bounded context assembly, mandatory authoritative-rule handling, supersession suppression, temporal and continuation signals, outcome evidence and selection-only checkout telemetry.
Broader cross-segment graph traversal, longer-horizon consolidation policy, richer outcome learning and comparative public benchmark coverage remain active development areas.
The first AMES result is a synthetic contract verification suite. It confirms intended behaviours but does not by itself prove superiority over another memory architecture.
Protected implementation
Speedway discloses the cognitive objective and evidence—not the recipe required to clone it.
Exact equations, weights, thresholds, tie-breakers, private prompts, routing recipes, graph traversal policy, consolidation triggers, internal schemas, private evaluation corpora and security-sensitive topology are intentionally withheld as proprietary or security-sensitive intellectual property.
Signal families, architecture, observable behaviour, version status, benchmark methodology, sanitized evidence, limitations and known failure classes.
The enabling implementation details that would permit direct replication, reverse engineering, attack optimization or appropriation of Speedway's proprietary methods.
Our transparency standard is that material technical claims should become measurable and versioned, not that the implementation itself must be disclosed.
CMI v2