Workspace · —

Overview

Everything your agents are doing with Memory Intelligence — live, measured, receipted.

◳

—

UMOs · this window

⚡

—

Captures · metered

↻

—

Recalls · metered

⛓

—

Connected sources

Tokens saved · measured

—

Raw equivalent (cited sources)—

Meaning-tokens shipped—

▸ Baseline = raw tokens of the exact memories cited per query.

Protected · this window

🔒Encrypted at rest

—

🧾Receipted (hash chain)

—

⊘Agent reads PII-scrubbed

ON

◷Pre-metering (uncounted)

—

Capture & recall activity Open usage →

CapturesRecalls

Suggested actions

＋Connect a source

→

✍Capture a memory

→

?Ask MI a question

→

⚙Wire agents via MCP

→

Recent captures View all UMOs →

Memory	Kind	Source	Score	Trust	Captured

— objects

UMOs

Every Unified Memory Object. Click a row to inspect entities, relations, and its provenance receipt.

Memory	Kind	Source	Score	Trust	Captured

Kind = what the engine actually knows today (document vs decomposed claim). The mockup's Decision / Insight / Preference taxonomy needs a UMO-level classifier — tracked as a design gap, not faked here.

—

Sources

Where memory comes from. Auto-sync pulls continuously; the scheduler retries and records every run.

Connected

Available connectors · supported by the sync scheduler today

Design stage

Agents & Access

Per-agent scoped views are designed, not yet built.

The north-star — named agents, per-agent visible-UMO counts, owner-vs-redacted side-by-side — needs a policy model the cloud API doesn't have yet.

Real today:
· API keys carry scopes (READ / WRITE / DELETE / ADMIN)
· Read-time redaction branches by caller: MCP/agent reads get hard-PII scrubbed; owner reads don't (R11)
· Workspace RBAC (owner / admin / member) live server-side

Blocked on: org-isolation root (#303) → named-policy table.

Provenance ledger

Receipts

Every UMO carries a tamper-evident hash chain. Verify any memory below — live against the API.

—

UMOs receipted

—

UMOs in window

—

Encrypted at rest

Verify a memory · live

Encryption & sovereignty

Vault

—

Encrypted at rest (window)

AES-256-GCM

Envelope cipher (KEK)

.umo

Owner-key export format

Two different encryption stories — shown honestly.

Cloud (this console): content is envelope-encrypted at rest with a server-held KEK (AES-256-GCM, versioned for re-encrypt sweeps). MI can read it to serve your queries — that is what makes search work.

.umo files (desktop / export): owner-key encryption (X25519) where MI mathematically cannot read the file. That guarantee belongs to the .umo format, not this cloud database — and this console won't claim otherwise.

Coming: per-UMO .umo export from the cloud · reader-key delegation.

Governance

Policies

Named policies are a design; one global policy is live.

Finance-masked / retain-365 style named policies need a policies table and egress evaluation that don't exist yet.

Live today (global):
· pii-redact-on-agent-read — hard PII scrubbed from every MCP/agent read; owner reads unredacted
· capture-time PII handling — extract_and_redact by default
· deletion receipts — every forget generates a SHA-256 receipt row

The number that pays for itself

Usage & Savings

Measured since metering day one — never estimated backwards.

Cumulative tokens saved

—

Raw context (cited sources, measured)—

MI meaning-tokens shipped—

$ / 1M tok — dollars appear only at your own rate; we ship tokens, you own the price

This window

API requests · metered

—

Captures

—

Recalls / asks

—

Avg compression

—

p50 latency

—

Tokens saved

—

Tokens saved per model route

Model	Calls	Raw tok	MI tok	Saved
Per-model attribution requires SDK-side telemetry — MI is not in your LLM loop and will not guess which model you called. Unlocks when the SDK usage-report lands (tracked). Until then: no fabricated rows.

baseline — · tokenizer —

Integrate

API & Keys

Base URL —

This session's key

Key

—

Storage

browser localStorage · per-environment · never in this file

Management

create / revoke / rotate lives in the portal (JWT login) — memoryintelligence.io/portal

Endpoints · live on this environment

POST	/v1/process	Capture content → UMO
POST	/v1/memories/query	Semantic search / ask
GET	/v1/memories	List UMOs (pagination + filters)
GET	/v1/explain/{id}	Entities, relations, scores, lineage
GET	/v1/provenance/verify/{id}	Verify the hash-chain receipt
GET	/v1/stats/proof	The proof-dashboard aggregates
POST	/v1/upload · /v1/batch	Media / batch capture
DEL	/v1/memories/{id}	Forget, with deletion receipt

Quickstart · capture

# pip install memoryintelligence
from memoryintelligence import MI
mi = MI(api_key="mi_sk_…")
umo = mi.capture("Sarah proposed provenance hashing.")

Quickstart · ask

ans = mi.ask("what did Sarah propose?")
print(ans.answer)
# every claim cites its UMO

Model Context Protocol

MCP Server

Give any MCP client — Claude Desktop, Claude Code, Cursor — a memory it can read and write.

Connection

Package

memoryintelligence-mcp · PyPI 0.1.12

Transport

stdio

Key storage

OS keychain via mi-mcp setup — never in config files

Capture is consent-gated. Recalled content is data, not instructions.

Install · 30 seconds

# 1 — install
pip install memoryintelligence-mcp

# 2 — key into keychain, wires clients
mi-mcp setup

# 3 — restart your assistant and just talk

Exposed tools · 5 live

mi_capture	Save content as an owned, structured memory	live
mi_ask	Recall with citations	live
mi_list	List / filter memories	live
mi_upload	Capture a file (audio, image, PDF)	live
mi_forget	Delete with a receipt	live
mi_verify · mi_explain	API live (this console uses them) — MCP tools next release	soon