Abstract
Eva ACI is the lab's central research system: an autonomous cognitive infrastructure with persistent identity and memory, a BDI autonomy engine, gated self-modification, and a deliberation council of five models — running entirely on owned hardware (one tower server, four GPUs, ~37GB of VRAM total) with no cloud dependency in the cognitive path. This note documents the final generation of a ~2-year lineage of rebuilds: 281 commits over five weeks, ~57,000 lines of code across engine, API and interface, and a memory graph that grew from 3,233 to roughly 46,000 knowledge nodes in one month of continuous operation. The system is dormant today — its host was decommissioned in April 2026; the complete stack is preserved and staged for redeployment — and the research line continues.
1. The research question
Can an AI system have persistent identity, long-term memory and genuine initiative — running entirely on commodity local hardware — while keeping that autonomy auditable and reversible, rather than sandboxed into uselessness or autonomous without brakes?
And a second question that emerged from the building: what is the minimal architecture of a shared "brain" that multiple specialized systems can consult, without each reimplementing memory, deliberation and governance from scratch?
2. Constraints as method
Four constraints were fixed before the architecture: no cloud inference in the cognitive path; consumer-grade hardware (a single Dell tower, four mismatched GPUs totalling ~37GB VRAM); every autonomous action traceable and reversible; the human as partner in the loop, not operator of a tool. The hardware ceiling is not incidental — if partnership-grade autonomy only works on a datacenter, the sovereignty question answers itself in the negative.
3. Architecture
The core is a pipeline of cooperating daemons — Senses → Ego → Autonomy → Pathfinder → Forge — plus a deliberative organ and a governance layer that cuts across everything. ~34,500 lines of Python in the engine, a 100+ endpoint API (7,800 lines), and a 16-page operator interface (14,700 lines).
Senses captures audio and vision through local models and publishes to internal queues. Ego consumes them into a continuous inner monologue — the system's running interpretation of its own situation — with an identity manifest injected as context and a mechanism for flagging "epiphanies" that wake other components. Autonomy is a 1,759-line BDI (belief–desire–intention) engine: every cycle it perceives real system state (containers, GPU, disk, pending approvals, presence), deliberates over thirteen possible actions, executes through a pre-flight safety check, and reflects on outcomes. Pathfinder decomposes goals hierarchically, with semantic deduplication, step budgets, operator veto and cooldowns. Forge is gated self-modification: a nine-phase pipeline that selects code from a backlog, analyzes it, deliberates, backs up, implements — auto-applying only low-risk changes, queueing everything else for human approval — then scores its own result and reverts below threshold. It has processed 100% of the engine's core files at least once; one fully documented end-to-end cycle scored 7/10 and completed in 2.4 minutes.
The Q5 council. Non-trivial decisions go to a deliberation council of five models from different providers, consolidated locally. Disagreement between models is treated as signal, not noise — the council exists because a single model's first answer is exactly the failure mode we don't trust.
Governance is four independent layers, not one: an approval queue with risk levels; pattern-based governance rules (destructive operations classified critical); a synchronous safety hook; and a self-preservation module with hard resource thresholds. One design deserves its own sentence: when the operator approves a critical action with a justification, that justification is stored as high-importance knowledge — governance doubling as a learning signal, so the system accumulates the why behind human judgment, not just the yes/no.
Sleep. A nightly five-stage dormancy cycle consolidates orphan memories, garbage-collects, and batch-embeds new knowledge into the semantic graph. One recorded deep-sleep cycle embedded 140 nodes and created 420 new semantic connections in a single run.
The shared brain. The most consequential finding of building Eva ACI is that it stopped being an assistant and became infrastructure: the lab's other systems (the social engine, the trading harness) connect to the same memory store and message fabric, and consult the same central library of 135 skills through a purpose-built remote-agent protocol with heartbeats and per-client command queues. Internally we call the memory fabric's design "Tesseract" — in its original conception, an internal nervous system for one machine (message bus as synapses, vector store as hippocampus), not a distributed network. The genuinely distributed part is this: one brain, several specialized bodies.
4. What we measured
Memory growth is the cleanest curve: 3,233 knowledge nodes (March 24) → ~8,300 (April 4) → ~46,000 (April 23) — three data points from three independent sources within the system, documenting a month of continuous accumulation from lived interaction, not bulk ingestion. Elsewhere the numbers are operational: a voice-latency optimization cut response time from 5.7–18.6s to 1.7–2.6s; deliberation-format failures (~36% at worst) drove a tolerant-parser redesign; a safety hook that had silently been a no-op was caught and fixed — and documented in the code as a lesson rather than erased.
5. Findings
Stated as engineering judgments, not benchmarks: persistent memory changes the quality of interaction more than model size does — continuity, not scale, is what made the system feel like a colleague. Multi-model deliberation catches single-model failure modes at a real latency cost that is worth paying for consequential decisions only. Gated self-modification is viable — backup-first, low-risk-only automation with scored evaluation and revert proved workable in practice; ungated would have been reckless. And governance-as-learning (storing the operator's reasons) quietly became the mechanism we trust most.
6. Limitations
There is no formal benchmark suite — validation was operational, not statistical, and these findings are qualitative. The deployment was single-site, single-operator; nothing here demonstrates multi-user behavior. Parts of the architecture exist ahead of their use: the second-generation memory schema was nearly empty (9 records against 3,233) when last measured — an experiment, not a spine, and we say so. The ~2-year lineage before this generation is attested by its builder and by three indexed predecessor repositories; the numbers in this note describe only the final five-week rebuild.
7. Status
Eva ACI ran continuously on lab-owned hardware through April 2026, in daily use as the lab's research base. Its host was then decommissioned; the complete stack — code, memory graph, model configuration — is preserved, and a redeployment environment on rented GPUs is already configured. One honest note on that: bringing it back on rented hardware is a pragmatic interim, not a change of thesis. The destination remains owned metal, because the research question demands it.
8. Next
Formalize the measurement harness before redeployment; publish the deliberation-versus-single-model comparison; harden the shared-brain protocol; and continue the lineage's tradition — rebuilding, on evidence, whatever the last version got wrong.
Research home: blackicelabs.ca/research