A first-class external memory you can plug into any AI. Private, local, $0 — it remembers across sessions, and shares one brain with your other AI tools over MCP.
🚧 In active development since 10 April 2026. Interested? Join us at https://discord.gg/RzXcvsabKD
This is the public information & GitHub Pages repository for TerranSoul — what it is, why it exists, and the research + benchmarks behind it.
TerranSoul keeps your language model frozen and makes the memory around it smarter — so a small local model behaves above its weight class, every change is auditable, and your knowledge stays portable data, not opaque weights. One brain, shared over MCP with your other AI tools.
📖 TerranSoul in Six Scenes — a non-technical primer · 📄 Research — From Memo to Memory — the thesis, with receipts.
Most AI assistants forget everything when you close the tab, send your data to the cloud on every call, and charge a monthly subscription for it.
TerranSoul runs on your own machine: local models on your own hardware for $0, data that never leaves your device, and persistent offline memory that picks up where you left off. It shares one brain over MCP, so your other AI tools (Claude Code, Copilot, Cursor, Codex, peer instances) get smarter too.
TerranSoul is a first-class external memory — built for knowledge representation, retrieval, and updating, the way a brain organizes and revises what it knows: a typed knowledge graph, three decaying memory tiers, hybrid retrieval, and a closed write → manage → read loop. The language model stays frozen; the memory does the learning, so gains are auditable and reversible because they live in data, not model weights.
That makes TerranSoul a different kind of tool than Claude Code, OpenClaw, or SIA — none of those are memory-first. They are coding agents and self-improvement loops; TerranSoul is the memory substrate they plug into over MCP, and its sweet spot is memory-heavy read + update: long-lived, cross-session, cross-tool knowledge that has to stay correct as it changes. The benchmarks below show where a memory-first system lands — highest on the memory axis it was built for, and competitive even on tasks designed for the others.
Measured, chart-by-chart: a memory / retrieval head-to-head vs ~20 systems (
COMPARISON.md) and SIA's own benchmark suite run on a permanently frozen model — memory + iteration only, no weight training (SELF-IMPROVE-COMPARISON.md). The only estimate is the labeled H100 projection.
Memory recall — LongMemEval-S (500 questions): R@5 98.6% · R@10 99.8% · R@20 100% — the top score in the table (agentmemory 95.2%, MemoryPalace ~96.6%). SIA ships no memory subsystem.
Personal-assistant answer quality (22 prompts, independent 0–10 judge, same model): TerranSoul 9.82 at ~1.0 s for $0 — the highest and fastest local row.
Frozen-model self-improvement — ZorkGPT (4 B model): memory alone lifts the score from 0 (both AGI-pure controls) to 10–20, reaching rooms it never visited. No weight edits.
LawBench (191-class charge prediction): frozen 12 B + memory 76.3% Top-1, above SIA's weight-trained 120 B (70.1%) and prior SOTA (45%); DeepSeek-v4-pro frozen reaches 80.0%.
AlphaFold-3 TriMul kernel: a frozen agent reaches 3.87× over fp32 (RTX 3080 Ti) → ~14–15× H100-estimated ≈ SIA's 14×. The H100 figure is an estimate (re-bench TODO).
scRNA-seq denoising: a real frozen-actor denoiser on PBMC3k, raw MSE 0.046 (+35%). A raw scale, not SIA's normalized 0.289 — indicative, not a head-to-head bar.
OpenAI MLE-Bench Hard: SIA ranks #1; not run on TerranSoul — needs the MLE-Bench harness, Kaggle data, and multi-hour GPU runs (a truthful blocker, not a fabricated number).
Full methodology, per-task tables, and the 18-config embedder audit:
COMPARISON.md·SELF-IMPROVE-COMPARISON.md.
TerranSoul's thesis — memory, not weights — is defended in a research report, with receipts:
📄 From Memo to Memory: A First-Class External-Memory Architecture for Frozen Language Models — a combined position-and-measurement paper. It accepts the memo-vs-memory charge in the recent literature and contests the dichotomy: five behavioural criteria separate a memo from a memory, and a structured external substrate can satisfy them with the model's weights frozen — shown on a controlled Zork I study (ZorkGPT × external-memory bench) and four public benchmarks. (per-turn runs)
Darren Bui — darren.bui@terransoul.com
Interested in becoming a contributor? Join the Discord at https://discord.gg/RzXcvsabKD or email Darren directly. Devs, designers, VRM artists, prompt engineers, testers, and non-technical users are all welcome.
Built for the community. MIT License.