Raidho coding agent uses VSA memory instead of RAG
Raidho is an open-source alpha coding agent that replaces RAG with compositional Vector Symbolic Architecture (VSA) memory, splits reasoning and execution across different AI providers, and ships a reproducible benchmark showing a hybrid mode matched full tool-loop quality at ×2.6 less cost.
Score breakdown
Raidho's benchmark demonstrates that separating reasoning from execution across providers — combined with VSA memory instead of RAG — can match full tool-loop output quality at ×2.6 lower cost on the same task.
- 01Raidho is an alpha-stage, open-source coding agent licensed under AGPL-3.0.
- 02It uses compositional VSA (Vector Symbolic Architecture) memory that persists across runs, instead of RAG.
- 03Reasoning (text mode, no tools) and execution (agentic tool loop) can run on different AI providers.
Raidho (named after the runic character ᚱ, meaning "journey/movement") is an open-source, alpha-stage coding agent that departs from the standard single-model-in-a-tool-loop architecture. Instead, it splits work across two roles: a "smart, expensive model" handles reasoning and planning in text mode (no tools), while a "cheap, fast model" handles execution in code mode (agentic tool loop). These two roles can run on entirely different providers, allowing configurations such as planning on Claude and executing on DeepSeek.
Memory is handled through compositional VSA (Vector Symbolic Architecture) rather than RAG, and persists across runs.
Memory is handled through compositional VSA (Vector Symbolic Architecture) rather than RAG, and persists across runs. The project also features a "Council mode" in which two providers debate a question and a neutral pass distills points of agreement, residual disagreements, and a recommendation — described as depersonalized and provider-pluggable with no built-in personas. A reproducible real-API benchmark ships with the repository (`benchmarks/real_task_opus.py`), with full evidence in `evidence/2026-06-11_opus_vs_raidho/`. Running the same task against the same model, the benchmark recorded three cost points: a deterministic procedure at $0.05, a context-first hybrid at $0.116, and a pure tool-loop at $0.301, with the hybrid matching the tool-loop's report quality at ×2.6 less cost. The project is licensed under AGPL-3.0 and notes that APIs may change before a 1.0 release.
Key facts
- 01Raidho is an alpha-stage, open-source coding agent licensed under AGPL-3.0.
- 02It uses compositional VSA (Vector Symbolic Architecture) memory that persists across runs, instead of RAG.
- 03Reasoning (text mode, no tools) and execution (agentic tool loop) can run on different AI providers.
- 04Tested end-to-end against DeepSeek and Claude via the official Anthropic SDK.
- 05A reproducible benchmark shows three cost tiers for the same task: $0.05 (deterministic), $0.116 (context-first hybrid), $0.301 (pure tool-loop).
- 06The hybrid mode matched the pure tool-loop's report quality at ×2.6 less cost.
- 07"Council mode" lets two providers debate a question, with a neutral pass distilling consensus, disagreements, and a recommendation.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 15, 2026 · 11:57 UTC. How this works →