alethech

alethech

Verifiable agent continuity protocol — local-first, zero-LLM, zero-blockchain.

The art of un-concealing transmission integrity.

Copyright (c) 2026 AliceLabs LLC.

Listed in [TeleAI-UAGI/Awesome-Agent-Memory](https://github.com/TeleAI-UAGI/Awesome-Agent-Memory) (Emerging projects) after a maintainer audit of the claims against the code.

Now in the [official MCP Registry](https://registry.modelcontextprotocol.io/): [`io.github.eddyflores100-lang/alethech`](https://registry.modelcontextprotocol.io/v0.1/servers/io.github.eddyflores100-lang%2Falethech/versions) — the read-only alethech-mcp bridge, discoverable by any registry-aware MCP client.

If alethech is useful to you, a ★ [star](https://github.com/eddyflores100-lang/alethech/stargazers) is how other people — and their agents — find it.

From Greek ἀλήθεια (aletheia, "truth as un-concealment") + τέχνη (techne, "art, craft"). The art of revealing that a memory was not modified after being signed.

This is the reference implementation of the protocol specified in docs/implementacion-nucleo-minimo.md (rev 2 + rev 3 identity layer).

what this is

alethech provides cryptographically verifiable, portable memory continuity for AI agents. It signs memory commits with Ed25519, links them in a hash-linked DAG, rotates keys without losing identity, and can seal a verified history into one encrypted .aleth file for transfer between compatible runtimes and devices.

What this repo IS:

• A cryptographic protocol implementation (Ed25519 + SHA-256 + JCS RFC 8785)
• An encrypted portable memory container (.aleth, scrypt + AES-256-GCM)
• A neutral verified adapter view for plugins and chat integrations
• A read-only MCP stdio bridge (python -m alethech.mcp_memory, or alethech-mcp after pip install alethech) exposing verified memory context and verification tools — credentials and paths stay configuration, never tool arguments
• 9 CLI commands (init, commit, evidence, verify, export, import, migrate, key rotate, key revoke)
• A test suite with mutation-guard paths (each guarantee has a test that fails when the check is defeated)
• MIT licensed, published on PyPI as alethech

What this repo is NOT:

• The protocol core is not a hosted retrieval platform; the portable skill adds local encrypted memory workflows and bounded lexical recall
• It is NOT the legacy memex project (Python/ChromaDB memory server)
• It has no Docker, no auto-update, no LLM calls, no required cloud
• It depends only on cryptography and click — no mem0ai, no chromadb, no ollama

The legacy memex codebase (167 commits, AliceLabs Proprietary License) is preserved in a separate repository: eddyflores100-lang/memex-legacy. It is not part of this repo and not installed by pip install alethech.

the property

This memory set forms part of a cryptographically verifiable history associated with a determined identity, whose commits can be independently verified with respect to their integrity, cryptographic authorship, and provenance relations.

When a reference checkpoint exists, it can additionally be verified that the presented history continues from that checkpoint.

The protocol does NOT prove that the agent's claims are true — only that they were signed by the identity that claims them. The truth being un-concealed is the truth about transmission integrity, not about content.

status

Released version 0.9.1 — portable encrypted memory (.aleth v2 with ALETH002 recovery across Python, TypeScript, and Rust) and chat capture into portable memory are live. CI runs the Python suite on 3.10–3.13 plus cross-language conformance and .aleth interoperability across Python, Rust, and TypeScript. No LLM, no required cloud, no blockchain, no consensus.

install the portable memory skill

npx skills add eddyflores100-lang/alethech --skill alethech --global

Ask your agent to save a checkpoint, resume a project, or carry memory to another AI/device. First save creates encrypted local memory automatically. No API key, account, LLM model download, server or MCP configuration is required. Python 3.10+ and a script-capable Agent Skills host are needed; missing protocol dependencies are installed once in a private environment. Later memory operations run offline. Optional cloud-folder backup copies ciphertext only and preserves changed remote revisions. Browser-only chats use the existing extension/offline viewer.

See portable skill setup and limits and the primary-source alternatives review.

install

pip install alethech

For development:

git clone https://github.com/eddyflores100-lang/alethech.git

cd alethech pip install -e ".[dev]"

usage

alethech init                              # generate identity + genesis commit

alethech commit --content # create signed MemoryCommit alethech evidence --tool # create signed EvidenceCommit --input --output alethech verify # verify the whole store, offline alethech export --output # portable package alethech import --input # import external memory alethech key rotate # rotate operational key alethech key revoke --key-id # revoke a key alethech migrate --to v0.2 # migrate identity layer

capture a chat, carry the file

Use the Alethech toolbar icon to capture the loaded conversation, review it, and create an encrypted .aleth with a new local identity and recovery code. Carry that file to another device, chat or IDE. The extension imports it and inserts selected context into an empty chat editor only on request; it never sends. For any text-capable client, export context.txt. MCP clients can use the local read-only memory connector. A single offline alethech.html also opens/creates files without installing an extension or registering an account.

See the complete capture and transfer guide and IDE/MCP setup. Capture requires only activeTab and scripting, with no global host or storage permission. Rendered chat capture cannot retrieve a provider's hidden memories or unloaded conversation history.

portable encrypted memory (.aleth)

Alethech 0.9 development adds a single encrypted file designed for drag-and-drop transfer:

from alethech import Alethech

agent = Alethech.initialize("./working-store") agent.commit({"fact": "portable memory"}) agent.seal("memory.aleth", "your-passphrase")

Plugin-style path: dropped file -> unlock -> verify -> neutral context

context = Alethech.drop_context("memory.aleth", "your-passphrase")

The .aleth container uses scrypt + AES-256-GCM. Its authenticated payload can be opened byte-exactly by the Python, Rust, and TypeScript implementations in CI. The portable file may carry the encrypted operational signing key so history can continue on another device, but it excludes root.key and recovery.key.

See docs/ALETH_CONTAINER_SPEC.md and docs/ADAPTER_CONTRACT.md.

what it does NOT do

• It does NOT prove that the agent's claims are true. Only that they were signed by the identity that claims them.
• It does NOT detect rollback without an external checkpoint.
• The local working store is NOT encrypted at rest; the portable .aleth container is encrypted.
• It does NOT delegate permissions between agents.
• It does NOT call any LLM.

what the name means

alethech comes from:

• aletheia (ἀλήθεια) — Greek for "truth", more precisely "un-concealment" (Heidegger's reading: truth as the act of revealing what was hidden)
• techne (τέχνη) — Greek for "art, craft, technique"

So alethech = "the art of un-concealing". The protocol un-conceals:

• whether a memory was modified after being signed
• who signed it
• what signed timestamp was recorded (NOT physical signing time — only the timestamp recorded in the artifact)
• what its provenance is
• whether the chain of custody is intact

It does NOT un-conceal whether the content is true — that's the agent's responsibility, not the protocol's.

tests

python -m pytest tests/

The Python test suite covers:

• crypto primitives (Ed25519, SHA-256, base64url, base32)
• JCS canonicalization (RFC 8785) — 56 conformance vectors including:
• integer/float serialization per ECMAScript Number.prototype.toString()
• -0 serializes as "0" (per RFC 8785 erratum, NOT preserved)
• scientific notation format (positive exponents keep '+', negative strip leading zeros)
• decimal vs scientific threshold (1e21 / 1e-6)
• UTF-16 key ordering with surrogate pairs
• string escaping (control chars, non-ASCII preservation)
• NaN/Infinity rejection
• agent_id derivation (cryptographic binding to public_key)
• MemoryCommit signing, tamper detection, wrong-key rejection
• EvidenceCommit signing (supports both legacy and V2 identity)
• all 9 CLI commands (init, commit, evidence, verify, export, import, migrate, key rotate, key revoke)
• checkpoint emission + rollback detection + causal continuity (ancestry check)
• identity_mismatch detection
• export/import roundtrip preservation
• tampered manifest rejection
• key rotation with cutoff_head reachability guarantee
• ancestry_check mutation guard (3 code-level defeats)
• recall-seam defeats (2 data-level mutations: active-keys tampering, cutoff_head mutation)
• root_id binding (RootAuthority ↔ IdentityRecord ↔ ControlEvent)
• checkpoint continuity (checkpoint HEAD must be ancestor of current HEAD)
• IdentityRecordV2 signature verification (against root public key)
• import causal continuity (ancestry check before writing)
• import hardening (symlinks, path traversal, archive bombs, artifact hash verification)

repository structure

alethech/                    # this repo — cryptographic protocol only

├── alethech/ # source: crypto, objects, store, verify, cli, canonical │ └── integrations/ # langchain + memex hooks ├── alethech-rs/ # Rust SDK (31 tests, cross-validation) ├── alethech-ts/ # TypeScript SDK (15 tests, Web Crypto API) ├── tests/ # Python behavioral, adversarial, mutation and conformance tests ├── conformance/ # 15 adversarial conformance vectors ├── docs/ # protocol specification ├── examples/ # basic_usage.py + memory_writeback.sh (ops pattern) ├── INTEGRATION.md # how to integrate with memex via MemexAlethechHook ├── INTEGRATION-VIGIL.md # fail-closed memory verification for watchdogs ├── CHANGELOG.