theoAI
THE MEMORY CONSOLE

Every customer becomes a living memory card.

Source-tagged facts, open loops, and agent learnings, each with provenance, freshness, and a confidence score. Every account gets a Context Quality Score you can inspect, edit, export, or delete at any time. It's yours.

4
Memory layers
5
Quality sub-scores
0
Training on your data
∞
Exportable
MEMORY CONSOLE, PREVIEW

Every customer becomes a living memory card.

Theo turns every conversation into governed, inspectable business context. Open loops. Agent learnings. Provenance on every fact. Editable, exportable, deletable, yours.

theo / memory → customers → irena r.CONTEXT QUALITY 82%
Irena Riva
Account #A-2049 • Spans 3 channels • 14 conversations
KNOWN FACTS4 · source-tagged
Household size
Family of four (spouse + two children)
96%Verified
Conversation , May 02•Last seen 12 days ago
Preferred channel
SMS over phone calls
88%
Agent meta , inferred•Last seen 3 days ago
Active policy
ICHRA , carrier renewal 2026-01-01
99%Human-approved
Uploaded PDF , May 09•Last seen Today
Recent interest
Dependent coverage rules
82%
Conversation , May 14•Last seen 1 day ago
AGENT META, HOW THEO SHOULD BEHAVE
Tone preference
Plain Spanish, avoid insurance jargon
Format preference
Bullet summaries before paragraph answers
Hard rule
Never recommend plans without licensed-agent escalation
Hard floor
OPEN LOOPS, SURFACE PROACTIVELY

Decisions, questions, and promises Theo will surface unprompted in the next conversation, even months later.

question
Customer asked whether dependents are covered under ICHRA
Awaiting licensed agent review
promise
Broker promised follow-up before open enrollment
Due 2025-11-15
decision
Switch carriers vs renew , not ready until OE
Revisit week of 2025-11-01
YOUR DATA
  • • Exportable on demand
  • • Deletable, destructive and final
  • • Never used to train any model
Tour the Memory Console →Read the MCIR 2.0 spec

FOUR MEMORY LAYERS

Not a vector blob. A real graph.

The memory graph is structured. Each layer has its own shape, retention, and retrieval rules, all governed by the MCIR protocol.

Episodic

Conversation turns + the artifacts they produced. Timeline-indexed so phrases like "the PDF from last week" resolve to a real range.

e.g. Irena uploaded a policy PDF on May 09
Semantic

Distilled facts about customers, accounts, and the business. Source-tagged, confidence-scored, freshness-aware.

e.g. Household of four · prefers SMS · active ICHRA policy
Procedural

Learned workflows + business rules. Repeated patterns get promoted into durable procedural memory so AI Workers learn how things are done here.

e.g. Always offer two carriers before recommending a plan
Memory Chain

Conversation-level synthesis: {kind, title, summary, entities, outcome}. Open loops + agent meta memories ride here so a fresh thread feels like Theo already knows the user.

e.g. Open loop: broker promised follow-up before open enrollment

CONTEXT QUALITY SCORE

A live grade for every account's memory.

One score per customer or account. Five sub-scores tell your team exactly where the gaps are: missing fields, stale facts, contradictions, and unverified entries.

OVERALL
82
Irena Riva · Account #A-2049
Completeness78%
Freshness92%
Contradictions96%
Source quality84%
Human verification71%
! Gaps: renewal date · spouse DOB · preferred carrier. Flag for human verification.

CAPABILITIES

Memory you can see. Edit. Export. Delete.

Inspect

Walk every fact, open loop, and agent meta memory for any customer. See where each one came from, when it was last verified, and how confident Theo is.

Edit

Correct mistakes. Flag a fact as immutable. Approve a regulated entry. The agent layer respects your edits and never silently overwrites them.

Export

One endpoint returns the full memory graph for any account in a documented schema. Open schema, your data, no vendor lock-in.

Delete

Destructive and final. Removes the memory graph, knowledge files, and conversation history from primary stores and replicas. No shadow copies.

YOUR MEMORY, NEVER TRAINED ON

The memory graph is your asset. Not our training set.

We persist memory so your AI Workers improve over time. We never use any of it to train a model. Exportable on demand. Deletable on request. Documented schema.

Read the privacy stackRead the MCIR 2.0 spec

Give your AI Workers a memory you can audit.

Get your API key in 30 seconds. Memory ships on by default, governed from day one.

START FREE →BUILD AN AI WORKERMCIR 2.0 SPEC