Memory Fabric engineering preview
Candidate API, source-provided SDKs and qualification boundaries for governed PostgreSQL memory. Not generally available.
Engineering preview — not generally available
Supported candidate boundary
- Typed, internal, unverified memory with nonempty source references; a source reference records origin rather than proving a claim true.
- PostgreSQL transactions, version-checked updates, atomic mutation idempotency, source provenance, audit and quota receipts.
- Tenant/project/namespace eligibility before full-text retrieval. Recall uses lexical search by default. Candidate vector, hybrid and graph capabilities depend on separately configured adapters and qualification.
- Nullable UTC valid_from, valid_to and expires_at; current reads hide records outside their validity window. Updates replace the complete validity window; omitted fields become null.
- Soft forget checks the expected version and legal holds. Retained plaintext history and backups remain recoverable; this is not irreversible erasure.
- Owner-provisioned tenant operation and active-content budgets are required. A missing budget fails closed; consumption must not be presented as an approved invoice.
Identity, scopes and namespace grants
The host must resolve the existing MC-1 principal and selected project using trusted first-party authentication, then freshly check the exact operation and namespace grant. Tenant/project IDs, policy version and actor identity are server-owned, not customer JSON fields. Existing inference permissions do not automatically grant memory authority.
The Console Memory Fabric section is default-hidden unless the deployment enables it. A recognized deep link can remain unavailable. Selecting a namespace or seeing a button does not grant permission; the backend rechecks authorization, revocation and tenant isolation for every operation.
Core candidate paths
Paths are relative to an operator-provided qualified origin. Their implementation in source does not establish production availability. POST requests use application/json. GET record and inspection requests require namespace as a query parameter.
- POST /v1/memory/write — create; an update additionally requires memory_id and expected_version together.
- POST /v1/memory/search — bounded PostgreSQL full-text pages.
- POST /v1/memory/recall — uses bounded lexical search by default.
- POST /v1/memory/forget — expected-version soft deletion, erased:false.
- GET /v1/memory/{id} — current authorized record.
- GET /v1/memory/{id}/history — version pages with separate history authority.
- GET /v1/memory/{id}/provenance — source evidence pages with separate provenance authority.
- GET /v1/memory/audit — scoped, paginated metadata with separate audit authority.
- GET /v1/memory/usage — namespace snapshot with separate usage authority; not a final charge.
Retries, pagination and errors
Writes and forget require an Idempotency-Key of 1–128 printable non-space ASCII characters. Retry an uncertain mutation with the same key and payload; changed payload conflicts. Internal replays do not duplicate audit, usage or quota. The SDKs never automatically retry mutations.
Success is {data,request_id}; search, history, provenance and audit results contain paginated data and next_cursor. Pass an opaque cursor unchanged with the same identity, namespace and query. Page size defaults to 20 and is limited to 100. Cursors grant no authority and do not provide snapshot isolation.
JSON numbers use finite IEEE 754 binary64 values; integers must be between -9007199254740991 and 9007199254740991. Encode exact wider integers and arbitrary-precision decimals as strings; floating-point numbers do not promise lossless decimal precision.
Request JSON is bounded to 65,536 UTF-8 bytes, with a five-second body-read deadline; this is not a whole-operation deadline. The host and database need their own bounded timeouts. Errors return a stable error.code and server-generated request_id, with no internal credentials or SQL.
- 401: obtain valid existing MC-1 authentication; 403: request the required authorized scope/grant rather than bypassing policy.
- 409: idempotency payload conflict or stale version/policy conflict; inspect the original operation before retrying.
- 429 QUOTA_EXCEEDED: the tenant operation/content budget is exhausted.
- 503 QUOTA_UNAVAILABLE or MEMORY_UNAVAILABLE: deployment, storage or budget configuration is unavailable; availability must not be fabricated.
TypeScript / JavaScript source quickstart
Run from the Memory Fabric candidate source checkout, using its source-provided SDK. No npm package publication or live API is implied. An operator must provide MEMORY_FABRIC_ORIGIN as an approved HTTPS origin, MC1_PROJECT_API_KEY as an authorized existing key, and an approved namespace and tenant budget. The request shape is tested against the local HTTP/PostgreSQL candidate.
import { MemoryClient } from './sdk/typescript/index.mjs';
const memory = new MemoryClient({
baseUrl: process.env.MEMORY_FABRIC_ORIGIN,
apiKey: process.env.MC1_PROJECT_API_KEY,
timeoutMs: 10_000,
});
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');
const evidence = await memory.provenance(record.memory_id, 'approved');Python source quickstart
The synchronous Python client uses the standard library. Set PYTHONPATH to sdk/python in the candidate source checkout, and supply the same operator-approved origin, key and namespace. HTTPS is required by default and redirects are disabled; loopback HTTP is a separately enabled local-test mode, not a production configuration.
import os
from memory_fabric import MemoryClient
memory = MemoryClient(
os.environ['MEMORY_FABRIC_ORIGIN'],
os.environ['MC1_PROJECT_API_KEY'],
timeout=10,
)
record = memory.write({
'namespace': 'approved',
'memory_type': 'semantic',
'content': {'fact': 'Source-linked application knowledge'},
'source_reference': {'uri': 'urn:application:approved-document'},
}, idempotency_key='application-owned-stable-write-key')
found = memory.get(record['memory_id'], 'approved')
evidence = memory.provenance(record['memory_id'], 'approved')PostgreSQL and Supabase connector candidates
Source ingestion is explicit, one-way and bounded. Approve exact schema/table/column mappings and a restricted source role; dry run is the default. Source RLS and authorization must remain effective, and every accepted row separately passes target memory authorization, quota and provenance checks.
The connector does not automatically copy a database, enable CDC, synchronize deletes or write back. Primary-key continuation alone does not detect historical row updates or deletions. The Supabase candidate requires a verified source-identity adapter and its own source policies; local contract doubles do not prove a real issuer or application compatibility. Live PostgreSQL/Supabase demonstration, credentials, TLS and recovery remain qualification requirements.
CVoM and C³-R remain advisory
The deterministic compiler and conservative CVoM calculation estimate verified benefit minus lifecycle cost and risk. Uncertainty reduces benefit, insufficient calibration abstains, and hard security policy always overrides optimization. The module grants no commit authority.
C³-R integration remains research/shadow controlled. No empirical calibration or measured cost advantage is claimed here. Learned compaction, verified-fact revision, sharing and deletion require separate policy and qualification.
Qualification before commercial availability
- Verify existing identity, source permissions, revocation, quota and cross-tenant negative paths against the target environment.
- Qualify managed PostgreSQL concurrent behavior, network/TLS, backups, restore, migration rollback, load and outage recovery.
- Complete Console project/namespace onboarding and the real customer journey without preview adapters.
- Obtain approved prices and cost evidence, qualify durable billing reconciliation and payment lifecycle, and review tax configuration before enabling paid plans.
- Publish only after mandatory functional, security, commercial, website and operational gates pass. No Memory Fabric prices, SLA, certifications, dedicated regions or paper artifact are advertised by this preview.
