Home  /  Product

One layer, five operations.

Agents read through the Vault before they act, and write through it after. Everything below runs on the open-core package — it augments the stack you already have, rather than replacing it.

How it works

Five operations. Nothing more.

Open each one to see what the Vault does on your behalf.

Write a subject–predicate–object claim with optional validity bounds and visibility labels. The Vault classifies it against everything already known — it is never a silent overwrite.
Retrieve the claims that are current and authorized for the calling agent. Pass an as-of date to time-travel and resolve the belief state exactly as it stood on any past day.
Every incoming claim is classified by a six-way taxonomy — contradiction, supersession, refinement, duplicate, independent or ambiguous — so memory never silently corrupts.
Confidence decays over time and stale history compresses, so quality doesn't rot as volume grows. Decay verbs are themselves permission-gated and audited.
Every read and write lands on an append-only, hash-chained log. Replay reconstructs exactly what any agent could see at any moment — with a cryptographic receipt.
Human-in-the-loop

When the engine can't be sure, a human decides.

Reconcile resolves most conflicts automatically — by recency, source-trust, and confidence. Genuinely ambiguous cases, high-stakes writes, and calls that need a quality gate never get a silent guess: they escalate to a person, and the decision lands on the audit chain, attributed and tamper-evident. Human-in-the-loop is a configuration, not an afterthought.

  1. HeuristicLocalDeterministic rules resolve the clear cases instantly.
  2. LLM-judgeCloudAmbiguous writes are scored against everything already known.
  3. Human reviewEnterpriseNear-ties and high-stakes calls escalate to a person.
Review queue1 awaiting
cust:8841 · risk_rating
“elevated”new · KYC todayvs“low”Apr 2
near-tie · confidence 0.52 · the judge won't guess
ApproveEditOverrideDefer
Every decision is logged to the audit chain — who, what, when.
For developersConfigure the escalation policy; near-ties interrupt to the review queue and surface on your endpoint — approve, reject, or override to resume.
For risk & complianceA named owner, an SLA, and a tamper-evident record of every human decision.
For your regulatorSatisfies human-oversight controls — EU AI Act Art. 14 and OCC SR 26-2.
Federation

Augment your stack. Don't migrate it.

Front your vector store

Qdrant today; Pinecone, Weaviate & pgvector adapters on the way — temporal, ACL & audit layered over your results.

no migration

Connect over MCP

One-click into Claude Code, Cursor and any MCP client — no agent rewiring.

no rewiring

Cross-LLM by design

Provider-neutral storage — Claude, GPT, Gemini, Llama or Mistral. Switch the model without migrating memory.

no lock-in
Built for

From a laptop to the enterprise.

Enterprises

Governance, compliance packs and an audit your examiner accepts — for teams putting agents on regulated data.

governance + compliance

Developers

One-line MCP install, Python & TypeScript SDKs, open core — drop governed memory into any agent.

open-core + SDKs

Run it locally

Self-host the free core in-memory or on SQLite — no auth, no cloud, full control on your own machine.

self-hosted, free
Developers

Install in one line.

quickstart.py
# connect over MCP — no rewiring
$ uvx --from context-vault-ai context-vault-mcp

from context_vault.sdk import VaultClient

client = VaultClient(vault, principal)

# read only what's current & authorized
ctx = client.resolve("risk for cust:8841")

# write — conflicts handled, never silent
client.assert_fact("cust:8841", "risk_rating", "elevated")
# → CONTRADICTION vs apr-2 → SUPERSEDED ✓
↳Quickstart5 minpyPython SDKreferencetsTypeScript SDKreferencemcpMCP serverguide

Bring an agent and a dataset.

Leave with an audit trail you can prove.

Request accessSee pricing