Research summary
Civilization Memory Architecture is a versioned system design for making an AI agent's retained experience inspectable rather than treating memory as an undifferentiated text store. The July 2026 preprint documents the path from Sun v0.00.01 through Rosette v0.00.06: a frozen-model runtime, four typed memory systems, rapid episodic binding and replay, approval-gated consolidation, trace-backed procedural memory, atomic context packets, and a bounded multi-scale graph view.
The central claim is structural. A system should be able to state what a remembered item is, where it came from, which policy admitted it, which competing evidence exists, and whether it reached the model path. The accompanying tests establish executable contracts and controlled integration boundaries; they do not establish improved open-domain accuracy, autonomous lifelong learning, production reliability, or human-equivalent memory.
Why a memory architecture needs types
Persistent context changes an agent from a stateless predictor into a system whose earlier observations can affect later actions. A flat retrieval store can recall text, but it does not distinguish a temporary working value from a witnessed event, an approved generalization, an executable procedure, or a policy decision. That ambiguity makes it difficult to answer basic operational questions: which source supports a claim, when was it approved, what happens when sources conflict, and which retrieved records actually influenced a model request?
The architecture therefore separates four memory systems: working memory for short-lived active state, episodic memory for ordered event anchors, semantic memory for approved abstractions, and procedural memory for outcome-backed skills. Cells carry source provenance, summaries, confidence and importance signals, timestamps, lifecycle information, and typed metadata. Links record temporal, similarity, causal, task, procedure, replay, and conflict relationships. An append-only audit sequence records writes, reads, updates, expirations, consolidations, and conflict operations.
This is not a claim that software reproduces biological memory. Complementary fast and slow learning systems provide an engineering inspiration: bind specific events quickly, then form more durable general knowledge through a distinct and more conservative transition.
Five invariants for governed recall
The report specifies five implementation invariants. First, conflict handling preserves both pieces of evidence and adds links or decisions; it never silently overwrites a cell. Second, semantic and procedural memory derived from prior material retain source identifiers and an explicit approval identifier. Third, inactive working cells are excluded from default retrieval. Fourth, control decisions are not serialized as though they were factual evidence. Fifth, snapshot loading fails closed when its envelope, canonical hash, topology, or referenced cells do not agree.
These constraints make mutation auditable. A consolidation proposal may be inspected, rejected, or regenerated without changing authoritative memory. Only a guarded transition with compatible types, present sources, complete evidence, policy authorization, and deterministic identity may create approved derived state. A rejected transition can emit a failure artifact, but it may not publish a partial schema, skill, context packet, or recovered graph.
A bounded six-version path
Each version adds one bounded responsibility while preserving the authority of existing factual evidence.
Sun: frozen runtime and intervention surface
Sun v0.00.01 supplies the compatibility base: a frozen Qwen runtime and a narrow Adapter-facing context-vector surface. Keeping the base model frozen isolates the memory-path experiment from weight drift. It does not make the architecture a claim about optimal adaptation; adapters, LoRA, and prefix tuning remain relevant alternatives.
Orion: governed multi-system memory
Orion v0.00.02 introduces typed cells, links, traces, deterministic retrieval, TTL behavior, and session-versus-global boundaries. Its lexical scoring baseline is intentionally transparent rather than learned. Stable identifiers break ties, allowing a later audit to reconstruct why one cell was selected over another.
Trifid: episode binding and replay
Trifid v0.00.03 binds scenes, ordered source cells, entities, goals, actions, outcomes, cues, and time intervals into episode frames. A future consolidation candidate is admissible only after complete replay evidence and a shared signature across the configured number of frames. Binding organizes evidence; it does not itself turn observations into durable facts.
Lagoon and Eagle: approved semantic and procedural change
Lagoon v0.00.04 forms semantic candidates from episodes only through an approval-bearing transition. Eagle v0.00.05 derives procedural candidates from execution traces while retaining both success and failure provenance. Repeated approved writes use stable fingerprints for idempotence. Opposing outcomes are represented as conflict relations, not as a license to erase an inconvenient predecessor.
Rosette: atomic context and multi-scale graphs
Rosette v0.00.06 constructs typed context packets and a honeycomb-style graph projection over episode, schema, and control scales. The packet and graph are query-specific views, not new authorities. Context budgets operate on atomic fact groups so that a preserve-both conflict pair is not split into a misleading partial statement. Traversal is deterministic and bounded by explicit type, scale, and budget filters.
From stored evidence to an audit trail
The important boundary is between evidence, policy, and view. An episodic cell records that an event entered the system. A semantic cell is an approved abstraction over evidence. A procedural cell is an approved strategy with an outcome trace. A conflict decision governs handling but does not certify a fact. A context packet or graph view describes a query-time selection and can be rebuilt without mutating the store.
With these distinctions, an audit can reconstruct the selected cells, the policy that admitted or excluded them, their source lineage, the decision that changed an Adapter-facing path, and the snapshot generation used for the request. Auditability is therefore an operational property of the write, read, and recovery paths - not a visualization attached after inference.
What the evaluation actually validates
The preprint reports a fresh WSL/CUDA reproduction of 40 Orion, 20 Trifid, and 29 Lagoon-Eagle-Rosette tests. A generated Stage150 artifact tree contained 153 files and passed seven top-level release gates. A controlled five-node graph fixture exercised three scales, three adjacency types, and deterministic traversal to every node. These checks provide regression and artifact-coverage evidence for the stated contracts.
The frozen-model diagnostic encoded one semantic cell as a finite 1 x 1 x 1024 vector with a true mask. Under the full control, the traced Adapter locations showed nonzero memory-delta norms of 7.7929 and 4.9193; the no-memory intervention produced zero at both locations. The diagnostic also reported no trainable Qwen parameters, no Qwen gradients, and unchanged weights. This demonstrates a controlled intervention boundary, not a task-accuracy result.
Failure injection is equally important. The documented cases include working-memory TTL expiration, incomplete replay, missing approval identifiers, repeated fingerprints, opposite outcomes, tight context budgets, mutated snapshots, and missing graph cells. Required responses are rejection, preservation, or fail-closed recovery - never silent repair that obscures the evidence trail.
Limits and disclosure boundary
The evidence is deliberately narrower than a general intelligence claim. Retrieval currently uses lexical overlap and hand-set coefficients, not a learned retriever or vector database. The fixtures are small and deterministic; they do not measure long-horizon user work, open-domain retrieval, latency at scale, multi-model transfer, concurrent mutation, or production service objectives. Approval identifiers are software preconditions, not a substitute for organizational authorization, authentication, or compliance controls.
The public package exposes responsibilities, transitions, negative controls, aggregate results, and sanitized manifests. It omits credentials, private payloads, host locations, deployment tokens, model files, and operational configuration. That boundary enables technical scrutiny while avoiding publication of deployment-sensitive material.
Conclusion
Civilization Memory Architecture treats long-lived AI memory as a governed system of typed evidence and guarded transitions. Rapid binding preserves sources; replay precedes candidate formation; approval gates semantic and procedural writes; conflicts preserve both sides; control state stays separate from facts; context groups remain atomic; and corrupted snapshots fail closed. The reported artifacts establish a reproducible engineering foundation for future memory-model training and evaluation. Demonstrating broad task gains, learned consolidation, operational scale, and robust transfer remains future work.