An exploration into theoretical agentic memory architecture—a learning computer that evolves its own neural network-like memory structures rather than serving as a static assistant.
Current version: v6.1.0 (see docs/META_LEDGER.md)
EvolveAI implements the Autopoietic Memory Theory, a novel approach to machine cognition that treats memory as a self-maintaining, self-organizing system. Unlike traditional databases or retrieval-augmented generation (RAG) systems, EvolveAI's memory actively evolves, decays, and restructures based on usage patterns.
| Principle | Description |
|---|---|
| Adaptive | Memory that learns, evolves, and restructures based on usage patterns |
| Accountable | Every memory operation is traceable with cryptographic audit trails |
| Retrievable | Intelligent recall mechanisms that surface relevant memories contextually |
This project is explicitly NOT:
- A chatbot or assistant
- Cloud-dependent (operates entirely locally)
- A static database (memory must evolve, decay, and restructure)
The project has been rewritten in Rust (v3.0 onward); the original TypeScript prototype is kept frozen for reference.
| Component | Location | State |
|---|---|---|
| Rust core — thermodynamic memory model: decay, trust scoring, zero-trust crystallization, hash-chained ledger | crates/evolve-core |
Active (the product) |
CLI — 9 commands over persistent state in ~/.evolve/memory.json |
crates/evolve-cli |
Active |
| Desktop shell — Tauri 2 app exposing the memory commands | src-tauri (frontend entry in ui/) |
Active; frontend build wiring incomplete (see REPO_REVIEW-2026-08-11) |
| Legacy TypeScript prototype (v1.0–v2.1) | src/, lib/, components/ |
Frozen since 2026-03-18; 164 vitest tests still pass; superseded by the Rust rewrite |
| Vendored dependency — GG-CORE runtime (git submodule) | vendor/GG-CORE |
Required path dependency |
The Cargo workspace has a path dependency on vendor/GG-CORE/core-runtime, so the submodule is required — a clone without it will not build.
git clone --recurse-submodules https://github.com/MythologIQ-Labs-LLC/EvolveAI.git
cd EvolveAI
# If you already cloned without submodules:
git submodule update --init --recursive# Run the test suites
cargo test -p evolve-core -p evolve-cli
# Run the CLI (state persists to ~/.evolve/memory.json)
cargo run -p evolve-cli -- help
cargo run -p evolve-cli -- add "Important fact to remember"
cargo run -p evolve-cli -- search "important fact"CLI commands:
| Command | Purpose |
|---|---|
add <content...> |
Store a memory, print its address |
search <query...> |
Find memories by relevance |
forget <address> |
Delete a memory by address |
feedback <address> |
Pin fibers (CrossReference event) |
dispute <address> [sev] |
Inject entropy (default severity 0.5) |
approve <address> |
Approve crystallization (L2→L3) |
profile |
Show cognitive profile summary |
slo |
Show SLO report |
ingest <file> |
Ingest a text file as memory chunks |
Building the Tauri shell on Linux requires the GTK/WebKit system libraries, e.g. on Debian/Ubuntu:
sudo apt-get install libwebkit2gtk-4.1-dev libgtk-3-dev librsvg2-dev \
build-essential curl wget file libssl-dev libayatana-appindicator3-devThen build the shell crate with cargo build -p evolve-app. Note: the frontend dev/build wiring referenced by tauri.conf.json is currently incomplete — see the repository review for status and roadmap.
The frozen v1–v2 prototype still runs its own test suite:
npm install
npm test # 164 vitest tests(npm run build / npm run lint are currently broken on the legacy side; see the repository review.)
The Neural Net Processor is the computational engine implementing the Autopoietic Memory Theory:
┌─────────────────────────────────────────────────────────────────┐
│ NEURAL NET PROCESSOR │
├─────────────────────────────────────────────────────────────────┤
│ ┌────────────────────────────────────────────────────────────┐ │
│ │ LIFECYCLE ORCHESTRATOR │ │
│ │ GROUNDING → SEMANTIC_PAUSE → ACTIVE_FLOW → DETACHMENT │ │
│ │ ↓ │ │
│ │ REM_SYNTHESIS │ │
│ └────────────────────────────────────────────────────────────┘ │
│ │ │
│ ┌──────────────────────────┴──────────────────────────────┐ │
│ │ PROCESSING CORE │ │
│ │ ┌─────────┐ ┌──────────┐ ┌─────────┐ │ │
│ │ │ ENCODER │ ←→ │ TIER MoE │ ←→ │ DECODER │ │ │
│ │ └─────────┘ └──────────┘ └─────────┘ │ │
│ │ │ │
│ │ ┌──────────────┐ ┌─────────────────────┐ │ │
│ │ │ DECAY ENGINE │ │ SHADOW GENOME │ │ │
│ │ │ (CMHL) │ │ INTERCEPTOR │ │ │
│ │ └──────────────┘ └─────────────────────┘ │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │ │
│ ════════════════════════════╪═══════════════════════════════ │
│ MEMORY BUS │
│ ════════════════════════════╪═══════════════════════════════ │
│ ┌────────────────────┼────────────────────┐ │
│ ▼ ▼ ▼ │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ L1 TRANSIENT│ │ L2 TEMPORAL │ │ L3 UOR │ │
│ │ CACHE │ │ GRAPH │ │ VAULT │ │
│ └─────────────┘ └─────────────┘ └─────────────┘ │
└─────────────────────────────────────────────────────────────────┘
- GROUNDING — Establish session context, load soul file, allocate fiber budget
- SEMANTIC_PAUSE — Safety check against Shadow Genome before execution
- ACTIVE_FLOW — Execute operations with full pipeline tracing
- DETACHMENT — Clear transient state, checkpoint L2 graph
- REM_SYNTHESIS — Learn from traces, crystallize stable memories to L3
| Tier | Name | Characteristics | Decay Rate |
|---|---|---|---|
| L1 | Transient Cache | Fast vector-based, TTL eviction | λ = 0.1 (aggressive) |
| L2 | Temporal Graph | CMHL decay, edge traversal, semantic relationships | λ = 0.001 (moderate) |
| L3 | UOR Vault | Immutable hash chain, O(1) lookup, cryptographic verification | λ = 0 (immortal) |
Memories are routed to tiers by the Memory Tier Score (MTS):
MTS = (S × Ws) + (A × Wa) + (P × Wp) - (C × Wc)
with S = sensitivity, A = accuracy requirement, P = privilege level, C = compute constraint. MTS > 0.8 → L3, MTS > 0.3 → L2, otherwise L1.
Decay is computed lazily on retrieval using exponential decay:
w_current = w₀ × e^(-λt)
where w₀ is the initial weight (salience at encoding), λ the tier-specific decay constant, and t the time since last access.
A negative-constraint immune system that blocks execution when intent matches known failure patterns:
- Tracks failure categories:
COMPLEXITY_VIOLATION,HALLUCINATION,SECURITY_REGRESSION, etc. - Uses cosine similarity against embedded failure traces
- Provides safety verdicts:
PASSorBLOCK
- Lifecycle: RELEASED (v6.1.0), governed by the QoreLogic A.E.G.I.S. meta-ledger (81 chained entries).
- Backlog: 7 of 14 items complete, 1 partial, 6 open — next up is BL-001 (deterministic policy gate), which unblocks BL-002/003/007. See docs/BACKLOG.md.
- Known gaps and roadmap: see docs/REPO_REVIEW-2026-08-11.md.
| Document | Description |
|---|---|
| CONCEPT.md | Project DNA — Why, Vibe, Anti-Goals |
| ARCHITECTURE_PLAN.md | Original architecture blueprint |
| AUTOPOIETIC_MEMORY_THEORY.md | Theoretical foundations |
| NEURAL_NET_PROCESSOR_DESIGN.md | Detailed design specification |
| SHADOW_GENOME.md | Failure-pattern immune system |
| PRISM_UOR_MDK_SUMMARY.md | UOR identity algebra summary |
| BACKLOG.md | Research-sourced backlog with status table |
| SYSTEM_STATE.md | Sealed system state snapshot |
| META_LEDGER.md | QoreLogic governance chain (do not hand-edit) |
| REPO_REVIEW-2026-08-11.md | Product-readiness review and roadmap |
docs/plan-*.md |
Per-release implementation plans (v2 → v6.1) |
docs/Research/ |
Comparative research artifacts |
This repository implements and experiments with memory metabolism — lifecycle orchestration, decay (CMHL), tier routing, and REM-synthesis consolidation — within the Agent Memory reference architecture. EvolveAI's Autopoietic Memory Theory remains its own theoretical framework; this alignment maps responsibilities, it does not merge doctrines.
Canonical doctrine this work maps to:
- Lifecycle state machine
- Forgetting, consolidation, and memory metabolism
- PAMA — mutation authority
- ADR-020 — governed uncertainty
- Conformance test plan
The governing boundary: EvolveAI's learned and heuristic signals may propose decay, retention, consolidation, or promotion. They do not self-authorize irreversible or canonical state changes — transition proposal remains separate from transition commit.
Implementation status: declared
Conformance evidence: none yet
This is an experimental research project. Contributions should align with the project's anti-goals and core principles.
- All changes must pass the Gate Tribunal audit (
/qor-audit— the QoreLogic command family isqor-*) - Code must adhere to Section 4 Razor constraints:
- Functions ≤ 40 lines
- Files ≤ 250 lines
- Nesting ≤ 3 levels
- All changes are logged in the META_LEDGER
Apache-2.0 — See LICENSE file
- Autopoietic Memory Theory — Foundational theoretical framework
- Prism UOR Foundation — Model Development Kit for identity algebra
- QoreLogic A.E.G.I.S. — Governance protocol for accountable development
Built with the QoreLogic A.E.G.I.S. lifecycle protocol