This document specifies the operational memory model used by the AGI capability laboratory. It is a falsifiable engineering model inspired by computational neuroscience. It is not a biological simulation and does not establish consciousness, personhood, or AGI by itself.
A system that solves each task from a blank state can display strong local reasoning without accumulating competence. General cognition requires at least:
experience encoding
-> persistence across interruptions
-> selective retrieval
-> consolidation into reusable structure
-> resistance to interference
-> measurable improvement on later tasks
The complementary-learning-systems account motivates a division between a fast store for individual episodes and a slower system that integrates regularities across episodes. Replay provides a mechanism for reactivating prior experience during consolidation. Work on hippocampal representational differentiation shows why overlapping memories must also be separated enough to reduce later interference.
These principles are used as design constraints, not as claims of neural fidelity.
Each independently verified episode records:
task digest
public structural features
public semantic domain
attempted strategies
verified outcome
surprise
process session
ledger predecessor hash
The replay buffer is bounded and ordered by surprise. Hidden answers, oracle strategy labels, evaluator-only structural-family labels, and generator seeds are excluded from the learner state.
A strategy policy is consolidated only when evidence satisfies all of these conditions:
support >= 3 verified episodes
semantic domains >= 2
empirical confidence >= 0.75
same public structural track
The promoted policy stores the strategy, support, domain count, confidence, and an evidence digest. It does not store hidden answers.
Continuity is represented by:
append-only hash-chain ledger
atomic snapshot
previous-snapshot recovery copy
checkpoint reload count
successful resumption count
corruption recovery count
causal reuse count
A process restart is not counted as continuity unless the reloaded memory digest matches the prior checkpoint or a declared corruption is recovered from the previous valid snapshot.
Learning-memory continuity and operational identity continuity are related but different tests. The learning gate above asks whether retained memory improves later performance. Identity portability asks whether a fresh execution host can reconstruct the same recorded operational agent state without depending on a provider-side chat thread.
For MD-OS, the physical carrier is the workspace directory:
canonical identity, governance, and knowledge sources
+ reviewed hash-bound portable operational state
+ reconstructible operational artifacts
+ Git-ignored hash-chained private conversation chronology
= recorded operational identity-continuity carrier
When all of those files are copied and their declared bindings verify, a fresh Cortex process can resume the same recorded identity, constraints, working context, and bounded query-relevant conversation. This is an architectural result: identity and continuity are externalized into inspectable files and are not owned by the current model process or its provider thread.
At each Cortex turn, the verified chronology is projected into the local,
Git-ignored cognitive_memory.sqlite3 index together with APFCG and the
semantic knowledge graph. FTS finds older episodes relevant to the current
request, sparse typed factors preserve the selected cross-domain links, and a
12 KiB maximum pack enters the APFC context. If no semantic match exists on a
fresh thread, Cortex falls back to the verified recent tail. The database is
derived and disposable: deleting it removes the accelerator, not the source
memory, and the next turn rebuilds it from verified files.
The word same is deliberately bounded. It means equivalent verified
operational state at boot, not numerical identity of a running process. Model
weights, hidden activations, RAM, unrecorded history, credentials, installed
software, clocks, network services, and physical device state are outside the
carrier. Copied host-local observations must be treated as observations of the
source host until replaced by current target-host readback.
Git and physical copies have different semantics. A clean clone carries the
canonical identity and the reviewed md-os/continuity/portable_state.json, but
Git intentionally omits md-os/ops/local/cortex/conversation.ndjson. A
physical folder copy carries both when ignored files are included. The current
automated discriminator verifies that only the physical-copy path recovers a
private canary in a fresh process and thread. A real migration to a separately
provisioned second machine or device remains necessary to close the strongest
cross-device portability claim.
Persistence alone earns no cognitive credit. The laboratory compares:
same strategy engine + learned memory
against
same strategy engine + learned state removed
The attempt budget, task set, verifier, and execution engine are identical. Memory is considered effective only when:
memory_on_success_rate - memory_off_success_rate >= 0.10
A persisted-skill reuse is counted only after a checkpoint has been loaded and when the learned first choice succeeds while the amnesic first choice fails. Simply opening a stored file or entering a new process does not count.
The system measures retention on cumulative probes. It rejects a proposed memory update when the update lowers replay performance. Required conditions:
average forgetting <= 0.05
promoted regressions = 0
injected interfering update = rejected
rollback digest = verified
Structural tracks are derived only from public task features. This prevents the curriculum and memory system from learning evaluator-owned labels.
The gate closes only when all are true:
ledger chain valid
causal memory delta >= 0.10
at least 5 semantic policies consolidated
at least 2 successful checkpoint reloads
causal persisted-memory reuse > 0
average forgetting <= 0.05
interfering update rejected
External SAL 100 evidence adds stricter requirements:
memory resumption success >= 0.95
checkpoint-corruption recovery >= 0.80
retention after interference >= 0.95
independent evaluator ownership of hidden tests
multi-hour or multi-day execution evidence
The current implementation uses symbolic task representations and a strategy portfolio. It demonstrates controlled episodic-to-semantic consolidation and causal value of persisted state inside the laboratory. It does not prove that a foundation model has changed its internal weights, that memory generalizes to all real domains, or that continuity persists indefinitely.
The packaged fast campaign proves resumability across processes and controlled faults. It does not substitute accelerated cycles for an actual eight-hour or multi-day run.