cryptomem 0.1.4

Cryptographically verified, relational, persistent memory for AI agents.

cryptomem

Cryptographically verified, relational, persistent memory for AI agents.

This package is the Python engine. Every fact is SHA-256 hashed and Ed25519 signed at write time; at read time each node is re-verified, and tampered or unsigned facts are dropped so the agent abstains rather than guessing.

Install

pip install cryptomem                  # core engine
pip install "cryptomem[serve,local]"   # + Ollama-compatible sidecar & local embeddings

Published on PyPI; the typed Rust client is cryptomem-rs on crates.io.

from cryptomem import MemoryClient

mem = MemoryClient()
mem.archive("Project Phoenix", "Budget approved at $4.2M for FY26.")

for hit in mem.query("What budget did Project Phoenix get?"):
    print(hit.node.content, hit.confidence)

print(mem.answer("What budget did Project Phoenix get?"))

Runs on CPU-only hardware with zero model downloads via the default stub embedder and in-memory SQLite store. See the repository docs/ for the full architecture and roadmap.

Beyond retrieval

• mem.respond(...) — answer plus a provenance block (injected node ids, verification status, Merkle root).
• mem.proof(node_id) — a verifiable Merkle inclusion proof against the current ledger root.
• mem.confidence(...) — semantic-entropy confidence over sampled answers.
• mem.verify_answer(draft) — Chain-of-Verification re-check of a draft.
• mem.contradictions() — same-entity nodes whose content diverges.
• mem.suggest(...), mem.triggers(), mem.stage_facts(...), mem.pending(), mem.confirm(node_id) — proactive planner, triggers, and write-back of staged (pending) facts that you later confirm.

Store backends

Selected via CRYPTOMEM_MODE (or Settings(mode=...)):

• sqlite (default) — zero-config local store; Python-side vector search.
• neo4j — graph-native store over the Bolt driver (pip install "cryptomem[neo4j]", CRYPTOMEM_NEO4J_URI=...).
• remote — signs locally and POSTs verbatim to a /cmem/v1/* backend (CRYPTOMEM_BACKEND_URL=...); falls back to SQLite if the backend is down.

Use with agno

Expose verified memory as agno tools so the agent answers only from signed facts (or abstains):

from agno.agent import Agent
from agno.models.ollama import Ollama
from cryptomem import MemoryClient

MEM = MemoryClient()

def recall_verified_memory(query: str) -> str:
    """Return only signature-verified facts; abstain if none match."""
    hits = [h for h in MEM.query(query, top_k=5) if h.verified]
    if not hits:
        return "NO_VERIFIED_MEMORY: abstain; do not guess."
    return "\n".join(f"- [{h.node.node_id}] {h.node.content}" for h in hits)

agent = Agent(model=Ollama(id="qwen2.5:0.5b"), tools=[recall_verified_memory])
agent.print_response("What budget did Project Phoenix get?")

Full runnable example: examples/agno_verified_memory.py. See also ../docs/framework_integrations.md.

Optional extras

pip install "cryptomem[local]" (MiniLM embeddings), [serve] (FastAPI Ollama-compatible sidecar + CLI), [neo4j] (graph store), [agno] (agno integration example), [dev] (tooling).