- Rust 96.4%
- Shell 1.7%
- Python 1.3%
- Dockerfile 0.2%
- JavaScript 0.2%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
934621dd38assemble_system_prompt's operator section (AkhConfig.system_prompt) fully replaced the psyche-derived identity body whenever it was set, so any akh with a configured system_prompt ran with no psyche/identity assembly at all — silently, since tool + convention sections still composed correctly and masked the loss. Both live workspaces on hephaistos (ptah, khnum) set one, so both were running psyche-less. Fix: the psyche-derived body is now always present and leads the assembly; a non-blank explicit prompt composes as an additional "# Operator Instructions" section immediately after it, rather than substituting for it. Ordering rationale: psyche establishes who the akh IS (derived from its actual state, machine-owned) and must never be silently dropped by an operator-authored prompt; the operator section — voice/persona/policy, human-owned — comes second so it still wins on tone conflicts. Several operator prompts open with their own "You are <Name>, ..." line that now sits next to the psyche body's own identity statement; that's accepted redundancy, not contradiction, and this function does not attempt to de-duplicate by parsing operator text. When `explicit` is absent or blank the output is unchanged — byte-identical to before this fix — so a workspace with no system_prompt (e.g. the cluster's thoth) sees no behavior change; a dedicated test asserts this directly against the psyche_body + tool_section + conventions composition. Updates the three tests that asserted the old replace semantics (assemble_composes_explicit_with_tool_section, load_agent_uses_explicit_prompt_from_config) and adds the byte-identical regression test. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> |
||
| .anima/wip | ||
| .cargo | ||
| .claude/skills/babysit-lane | ||
| .claude-plugin | ||
| .forgejo | ||
| .github/workflows | ||
| .solstice | ||
| agents | ||
| benches | ||
| benchmarks | ||
| book | ||
| charts/akh-medu | ||
| contrib | ||
| crates | ||
| data | ||
| docker | ||
| docs | ||
| eval_data | ||
| examples | ||
| hooks | ||
| packaging | ||
| plugin-skills | ||
| scripts | ||
| skills | ||
| src | ||
| tests | ||
| training | ||
| vendor | ||
| wit | ||
| .anima-owner | ||
| .dockerignore | ||
| .gitignore | ||
| .mcp.json | ||
| AGENTS.md | ||
| build.rs | ||
| Cargo.lock | ||
| Cargo.toml | ||
| CLAUDE.md | ||
| justfile | ||
| LICENSE | ||
| README.md | ||
| release.toml | ||
akh-medu
A neuro-symbolic AI engine made to run on the CPU. Made for people. Not for robber barons.
akh-medu combines hyperdimensional computing (Vector Symbolic Architecture) with knowledge graphs and symbolic reasoning. It runs entirely on the CPU with no LLM dependency, no GPU requirement, and no external NLP models.
- Infer new knowledge via spreading activation, backward chaining, and superposition reasoning
- Reason symbolically using e-graph rewrite rules (equality saturation)
- Search semantically using 10,000-bit binary hypervectors
- Act autonomously via an OODA-loop agent with 15 built-in tools
- Parse and generate natural language in 5 languages via a grammar framework
- Serve knowledge over REST and WebSocket APIs
Quick Start
cargo build --release
akh-medu init
akh-medu seed apply ontology
akh-medu grammar parse "Dogs are mammals" --ingest
akh-medu query --seeds Dog --depth 2
See the Quick Start Tutorial for a full walkthrough.
Architecture
┌──────────────────────────────────────────────────────────┐
│ Engine API │
├──────────┬──────────┬──────────┬──────────┬──────────────┤
│ VSA │Knowledge │Reasoning │Inference │ Agent │
│ Ops │ Graph │ (egg) │ Engine │ OODA Loop │
│ ─────────│──────────│──────────│──────────│──────────────│
│ HyperVec │petgraph │ rewrite │spreading │ 15 tools │
│ SIMD │oxigraph │ rules │backward │ planning │
│ ItemMem │SPARQL │ e-graphs │superpos. │ psyche │
├──────────┴──────────┴──────────┴──────────┴──────────────┤
│ Tiered Storage │
│ Hot (DashMap) · Warm (mmap) · Durable (redb) │
└──────────────────────────────────────────────────────────┘
The rest of the workspace
The engine is the root crate. Around it:
| Crate | Role |
|---|---|
seshat |
Shared corpus library — RAG over pgvector |
heka |
Postgres+pgvector evaluation store for long-term memory |
akhomed-store |
Postgres-backed durable store for relocatable Akhs |
anima-ai |
Multi-engine natural-language routing |
grug-data / grug-train |
Grug-Brain BabyLLM data generation and Burn training beachhead |
akh-egraph-bench |
E-graph reasoning microbenchmarks |
benchmarks/harness-core |
The agentic-task benchmark harness (below) |
Deployment: a Helm chart in charts/akh-medu, Linux and illumos
packaging in packaging/, and cargo binstall support that pulls prebuilt
akh binaries from the Forgejo releases rather than compiling.
Benchmarks
benchmarks/ is a reproducible harness for agentic tasks, not a marketing table. Each run
is pinned to a task-set hash and a harness SHA; once a task-set hash has a published
run against it, that hash is frozen — edits land under a new hash rather than silently
changing what an earlier row measured. benchmarks/LEADERBOARD.md is generated by
bench-tool leaderboard from the committed runs under benchmarks/results/, never edited
by hand.
The harnesses include illumos-e2e and illumos-orchestration — real operating-system
maintenance and fleet-orchestration tasks, graded autonomously on a VM — alongside
frontiercode, locomo and longmemeval. Published pass rates run from 100% on
fleet-orchestration down to 0% on frontiercode. They are recorded as measured.
Documentation
The full documentation is hosted at akh-medu.dev and covers:
- Installation -- build from source, feature flags
- Quick Start Tutorial -- first engine, symbols, queries, agent
- Concepts -- VSA, knowledge graphs, inference, reasoning
- Agent -- OODA loop, tools, planning, Jungian psyche
- Server -- REST and WebSocket APIs
- CLI Reference -- all commands and options
To build the docs locally:
cd book && mdbook serve
License
This repository is licensed under GPL-3.0-only. See the LICENSE file.
For integration into proprietary applications, contact the author — commercial licences are sold deliberately, and the money funds the FLOSS work.
Contributing
Contributions are welcome under the arrangement that pays for this project: contributors agree that their contribution may also be offered under a paid proprietary licence alongside the GPL. That is what lets proprietary vendors fund FLOSS work instead of free-riding on it. Open the discussion before a large change so nobody wastes an evening.
Dependencies and IP boundaries
- The engine ships no model weights and requires none. No LLM dependency, no GPU, no external NLP models — VSA operations, the knowledge graph, e-graph reasoning and the grammar framework are all implemented in this repository and run on the CPU.
- All third-party dependencies are unmodified upstream crates from crates.io under
their own licences, pinned by
Cargo.lock. Notable ones areegg(e-graphs),oxigraph(RDF/SPARQL),petgraph,redb,dashmapandburn. - No proprietary code or datasets are included, and no training corpus is committed.
data/holds hand-authored seed ontologies, compartment definitions and a schema.org schema pack;training/holds run configs and scripts only. Training data is generated bygrug-datain this repository. - Optional external services are the operator's own: PostgreSQL with pgvector for
seshatandheka, and any cloud model endpoint a benchmark harness is pointed at. akh-medu holds no credentials of its own and phones nothing home. - Everything an operator puts in — symbols, triples, corpora, benchmark runs — stays the operator's.