AXM  Protocol
Genesis kernel Core hub
Chat spoke Show spoke Embodied spoke
axm-genesis v1.2.0  ·  axm-chat v0.1.0  ·  Apache 2.0

Your AI conversations
as sovereign knowledge

Vector databases are search engines.
This is version control for knowledge.

Turn Claude, ChatGPT, and Gemini exports into cryptographically signed, offline-queryable knowledge shards. Every decision traced. Every claim provable. No cloud required.

Vendor APIs are memory black holes.

You spend hours reasoning through architecture, strategic planning, and complex decisions with LLMs. That cognitive work is trapped inside fragmented chat interfaces controlled by vendors.

When you need to remember why a specific path was chosen or what alternatives were rejected, the context is gone. Search fails because the decisions were never formally committed.

The graveyard of failed attempts is the most valuable part of the record. It prevents the same dead ends later.

The Protocol
Not a search index — a provenance record.

The AXM protocol was built to verify autonomous robotics doctrine offline. axm-chat applies that same cryptographic rigidity to your personal knowledge base. Every decision is signed, every claim is traceable, every shard is yours.

Pass 1 · No LLM
Import
One signed shard per conversation. BLAKE3 Merkle, ML-DSA-44.
Pass 2 · Local LLM
Distill
Episodic index, lens extraction, decision shard. Import runs first — without a model — so the shard is cryptographically sealed before any LLM touches it.
Query · DuckDB
Ask
Natural language to SQL. No LLM at query time.

Every shard is verifiable.

Each compiled shard contains the source document, structured claims, byte-level provenance, and a cryptographic signature. Tamper any byte and verification fails.

Full shard format specification on Genesis →

BLAKE3 Merkle Trees
Every claim and text block is hashed into a strict Merkle tree. Verify structural integrity of any shard entirely offline — no central server required.
ML-DSA-44 Signatures
Post-quantum signatures (FIPS 204) ensure that once a shard is compiled, the authorship and timestamp are mathematically immutable. Ed25519 also supported.
DuckDB + Parquet
Shards are stored locally as compressed Parquet files. At query time, the Spectra engine mounts them in-memory via DuckDB for instant relational analysis.
Lineage + References
When a decision shard supersedes a conversation shard, lineage@1 records the chain. references@1 links every decision claim back to the exact conversation turns that produced it.
Episodic Index
Every conversation gets a universal base pass: topic tags, named people, tools, project references, question text, resolution state, and tone — written to ext/episodes@1.parquet.
Engineering Lens
A gated second pass for technical conversations. Extracts the problem statement, graveyard of failed attempts, working solution, and any architectural rule — written to ext/engineering@1.parquet.
Install
pip install -e ./axm-genesis    # cryptographic compiler + verifier
pip install -e ./axm-core       # Spectra DuckDB query engine
pip install -e ./axm-chat       # chat spoke
Import
axm-chat import ./conversations.json    # one signed shard per conversation
Distill
ollama serve && ollama pull mistral
axm-chat distill                # episodic index + decision extraction
Query
axm-chat query "what decisions have we made"
axm-chat query "what failed before we solved the merkle problem"
axm-chat query "what changed since january"
axm-chat verify                 # offline Merkle + signature check

All data stays in ~/.axm/shards/ on your machine. Verify any shard entirely offline: axm-verify ./shard_dir

Memory has a shape.

AXM operates across three layers that share one container format but serve different temporal purposes. The protocol does not distinguish between them — all shards are structurally identical. The distinction is semantic, enforced by convention.