Persistent records
PostgreSQL-backed writes, optimistic versions and idempotent mutations. Local tests exercise the actual SQL engine; managed-service durability and recovery still require qualification.
Engineering preview — not generally available
A governed memory layer for applications and autonomous agents. Preserve source-linked knowledge, retrieve within authorized scopes, and make changes traceable across time.
This page describes a locally tested implementation candidate. Hosted provisioning, paid plans, production regions and operational qualification are not available through this preview.Persistent agent memory needs more than a similarity result. It needs an answer to who may read it, where it came from, when it is valid, and what happens when it changes.
Memory Fabric places that governance beside proven transactional storage. PostgreSQL supplies the database substrate; the memory layer supplies scoped operations, evidence and lifecycle controls.
The request path constructs an eligible namespace before searching. Model recommendations never grant access to records or authorize a commit.
Existing MC-1 principal; fresh operation and namespace checks.
Typed records, validity windows, version checks and retry handling.
Transactional records, isolated scopes and bounded full-text queries.
Source provenance, atomic audit receipts and quota accounting.
The current implementation is useful to inspect and qualify. These capabilities have local evidence; they are not a production availability or performance promise.
PostgreSQL-backed writes, optimistic versions and idempotent mutations. Local tests exercise the actual SQL engine; managed-service durability and recovery still require qualification.
Explicit validity windows and expiry govern current reads. A versioned update preserves the record of what changed; retention and legal holds remain policy boundaries.
Namespace authorization comes before SQL full-text search. Tenant, project and namespace restrictions remain inside the database boundary.
Source references, version history, audit events and namespace usage accompany operations. Owner-provisioned operation and content limits fail closed.
Records retain source references and version provenance. Identity and policy come from trusted server adapters, never from a tenant ID asserted in customer JSON.
Read current memory, page through history and provenance, and inspect scoped audit and usage records. A recorded source is evidence of origin; it does not automatically verify the source’s claim.
Content remains untrusted data. Instructions inside an ingested document cannot expand an agent’s authority.
The current endpoint performs version-checked soft deletion. Legal holds and policy can block it. Retained history and backups remain recoverable; this preview does not claim irreversible erasure.
Explore Agent Identity + Guard →CVoM is an advisory decision seam. Its conservative fallback works without an optional learned model, and insufficient calibration results in abstention.
Conservative uncertainty bounds reduce the estimated benefit. A hard policy denial wins; legal holds block economically motivated disposal. No learned controller is authorized to revise verified knowledge or delete a record.
C³-R integration remains a research and qualification boundary. This preview makes no empirical calibration, retrieval-quality or cost-saving claim.
Read the C³-R research boundary →PostgreSQL and Supabase connector candidates support incremental, permissioned ingestion. Live compatibility and the complete existing-application journey still require qualification.
Approve exact tables and columns, use a restricted source role, and begin with a bounded dry run. Source authorization and target memory authorization are separate checks. The connector does not copy an entire database or write back automatically.
The intended adoption path preserves existing Auth, PostgreSQL and row-level security. Verified source identity and policies must be enforced at the source and target. Local contract tests are not a real Supabase issuer or production compatibility receipt.
Source-provided TypeScript and Python clients implement explicit timeouts, structured errors and retry-safe mutation keys. They are preview source packages, not published production SDKs.
// From the Memory Fabric candidate source checkout.
// An operator must supply a qualified HTTPS origin and authorized project key.
import { MemoryClient } from './sdk/typescript/index.mjs';
const memory = new MemoryClient({
baseUrl: process.env.MEMORY_FABRIC_ORIGIN,
apiKey: process.env.MC1_PROJECT_API_KEY,
});
const record = await memory.write({
namespace: 'approved',
memory_type: 'semantic',
content: { fact: 'Source-linked application knowledge' },
source_reference: { uri: 'urn:application:approved-document' },
}, 'application-owned-stable-write-key');
const found = await memory.get(record.memory_id, 'approved');This shape is exercised against the local HTTP and PostgreSQL candidate. A deployment operator must provide a qualified origin, existing project identity, namespace grants and tenant budget before a real application can connect.
Read the candidate API and quickstart →Memory Fabric reuses established identity and control-plane seams. A related product’s availability does not prove Memory Fabric is production-ready.
Memory Fabric prices are not published or approved for purchase. Storage, operations, inference and verification costs require representative workloads and commercial review before a catalog is released.
Discuss residency, private networking, retention, workload size and support ownership. Available regions, dedicated deployments, certifications and SLAs remain subject to technical and contractual qualification.
Managed PostgreSQL recovery and load tests, end-to-end existing identity, source compatibility, billing lifecycle evidence and the complete customer journey must pass the release gate.
Vector retrieval, governed sharing, irreversible erasure and the wider Database/Auth/Storage/Realtime/Functions/Studio platform remain roadmap items, not live capabilities.
No self-service hosted release is available through this engineering preview. Contact ColomboAI to discuss requirements; the documentation describes the implementation candidate and its qualification gates.
Recall uses bounded PostgreSQL full-text search by default. Candidate vector and hybrid recall and graph retrieval require separate configuration and qualification. Hosted access is pending.
No. Current writes are internal and unverified. Source provenance records origin and change history; independent verification workflows remain a separate capability.
A Memory Fabric paper is not linked from this preview because no published artifact has been verified for this release. Existing C³-R research has its own explicit maturity boundary.
Review the candidate contract, define the authority boundary, and agree on the evidence required for a qualified deployment.