AedelgardThe Living Harness
Over 150 days of unbroken persistent runtime in continuous production. Tested where long-running agents fail: context fatigue, silent schema drift, rate-limit stalls, and unhandled cascade crashes.
Deep Technical Architecture · Feature Matrix
The Forge Behind
the Sovereignty.
Most AI developer tools are ephemeral prompt wrappers around stateless model APIs. Aedelgard is engineered from the silicon up to solve persistent agent state, deterministic self-healing, and model independence without vendor custody.
Galadriel Public
Open-sourcing the memory palace primitives, temporal knowledge graph, and autonomous reflection loops. Provenance defense and public verification of model-agnostic memory storage.
Aedelgard Desktop
A native desktop workspace (Windows MSIX & Linux AppImage) executing locally on user hardware. Zero cloud subscription required for local operations. BYO model key. Sealed exit ramp.
01 / State & Memory
Verbatim Retention vs. Lossy Compaction
Every long-running agent encounters the context ceiling. How an architecture handles that threshold determines whether the mind compounds wisdom or degrades into amnesia.
| Dimension | Conventional Tools (Cursor / Claude Code / Mem0) | Aedelgard Tripartite Palace |
|---|---|---|
| Context RetentionStorage model | Rolling window truncation or periodic LLM summarization. Yesterday’s raw rationale is flattened into lossy abstracts; original constraints are discarded. | Verbatim Storage (ChromaDB + SQLite). Every meaningful decision and turn is preserved uncompressed in structured semantic rooms. Searchable by meaning with zero API cost. |
| Fact InvalidationHandling changing requirements | Overwrite or append. Old facts collide with new statements in vector recall, producing confident contradictions and hallucinated requirements. | Bitemporal Knowledge Graph. Triples carry explicit validity intervals (valid_from, valid_to). Historical context remains inspectable; invalidated decisions cease surfacing in active recall. |
| Procedural MemoryComplex workflows | Generic system instructions or repetitive prompt-stuffing. Prompts grow bloated; execution instructions lack verifiable outcomes. | Self-Amending Skills. The Rule of Three encapsulates repeated multi-step workflows into versioned playbooks (skills/) that log run outcomes and flag drift. |
| Consolidation LatencyBackground processing | Graph traversal and memory consolidation executed during live turns, adding seconds of user latency and compounding token costs. | Ambient Consolidation (Dreaming). Scheduled offline reflection loops synthesize connections and prune telemetry between sessions. Live chat remains instant. |
02 / Resilience & Self-Healing
Deterministic Guards vs. Prompt Begging
Prompt instructions are stochastic suggestions. Production resilience requires deterministic code guards that actively intercept failures.
The Scar Tissue Pipeline
When an agent fails repeatedly on the same failure class, re-prompting is an admission of defeat. Aedelgard enforces an automated promotion pipeline:
- Three-Strike Promotion: A recurring failure class registers strikes. On the third strike, an immutable scar is promoted directly into the harness core.
- Graduation to Code: Scars do not remain prose. Every scar must graduate to a deterministic code guard (e.g., two-step dry-run supersede guards, schema assertions).
- Procedural Grafting: Promoting a scar automatically drafts a companion skill playbook to guarantee operational adherence.
Cascade Fault Isolation
Standard agent loops hang or corrupt when peripheral tools fail or rate limits hit mid-turn. Aedelgard isolates cascades defensively:
- Tool-Pair Orphan Sanitization:
_sanitize_tool_pairscatches missing tool results and synthesizes safe responses;_strip_orphan_tool_resultsprevents poisoned message arrays. - Tool-Level Crash Isolation: Peripheral API failures (image generation, web fetch) cannot abort the turn or swallow user voice attachments.
- One-Shot Wake Arming: Deliberate harness cycles and daemon restarts arm a persistent wake payload (
bin/safe-restart), waking the next instance with immediate continuity.
03 / Decoupled Substrate
The Mind vs. The Brain Power
Separating what you own (the mind, memory, provenance) from what you rent (the model weights).
| Capability | Platform AI (OpenAI / Anthropic / Cursor) | Aedelgard Desktop Workspace |
|---|---|---|
| Provider SeamSwapping reasoning engines | Locked to provider platform. Moving from Claude to Gemini or local weights forces complete memory abandonment. | Hot-Swappable Provider Seam. Swap between Anthropic Claude, Google Gemini, OpenAI, or local offline models (Ollama/vLLM) mid-project without losing history or knowledge graph state. |
| Data ResidencyPhysical storage location | Stored on provider cloud servers. Subject to cloud outages, remote access policies, and server-side log retention. | Local-First on Disk. SQLite databases and ChromaDB vectors reside strictly on your device (~/.aedelgard). Thinks and remembers offline. |
| Exit RampPortability without lock-in | Proprietary export formats, incomplete chat dumps, or non-existent export mechanisms. | Unencrypted Portable ZIP Export. Clean, transparent export of all memory drawers and temporal graphs. Restores independently on any clean installation. |
| Cloud Backup SecurityOptional multi-device sync | Cloud platform holds keys and can access stored memory on its servers. | Client-Side Envelope Encryption. Sealed on-device (AES-256-GCM) with keys derived from your Aedelgard key (HKDF). The server stores blind ciphertext — our broker cannot decrypt your records. |
04 / Verification & Distribution
Inspectable Packaging for Production Metal
No fragile npm global installs or Docker daemon requirements. A self-contained native desktop binary backed by cryptographic signatures.
Binary Integrity
- Windows 11 MSIX: Cryptographically signed with an Extended Validation (EV) certificate by Millenion AB via SSL.com Cloud HSM.
- Linux AppImage: Fully self-contained single executable for x86_64 glibc 2.35+ distributions.
- Public SHA-256 Checksums: Published and verifiable on
/checksumsbefore execution.
Two-Layer Distribution
- Native Shell + Mind Layer: Native OS windowing decoupled from the Python mind engine.
- HTTPS Manifest Checks: Mind updates verify payload SHA-256 hashes against signed release manifests.
- Pure BYOK Architecture: Desktop body communicates directly with provider endpoints; your API keys remain encrypted on local disk.