Our principles

A current record you can inspect.

Your agents need to know what the database currently accepts and how it got there. These seven principles guide how ConceptMem handles proposed updates, shared models, policy, and provenance.

A governed updateExample

Alice moved. The record can explain what changed.

Current state
Alice’s location is Lisbon.
History
Berlin is superseded, with its period of validity preserved.
Provenance
Profile update · account agent · confidence 0.98 · September 12, 2026 · structured input.
Policy
Location keeps one current value. The accepted update supersedes Berlin.
An illustrative update under a supersede policy. Repeating Lisbon leaves the current value unchanged; a later change keeps the earlier value in history.

Built into the workflow

Seven principles for accountable state.

01

Writes are proposals

An observation proposes a change. Consolidation decides what becomes current and reports applied, skipped, refused, or held updates. Schema changes follow a proposal path too.

02

Consolidate, do not append

Reconcile incoming information with existing facts. Repeated values leave the fact unchanged; conflicting values follow your policy. Keep a current record and the history of its changes.

03

Separate current state and history

Read accepted facts for the next action. Query earlier values, validity periods, and the changelog when you need to explain what changed and when.

04

Let the ontology govern

Share a model of entity types, fields, and relationships. Choose open or strict classes, or enforce one field’s type and allowed values. Stronger constraints are an explicit choice.

05

Require provenance

Record the source, actor, confidence, timestamp, and extraction method behind mutations. Make the origin of a change available to the people and agents who need to inspect it.

06

Treat projections as views

Generate Markdown pages, summaries, and exports from stored state. Submit a new observation to correct the record; edits to a derived view do not mutate the database.

07

Make policy pluggable

Configure update, conflict, and review rules with class defaults and field overrides. Retention and privacy settings can be stored and reported, but their enforcement is not implemented.

The service applies the rules across clients.

Connect through hosted REST or MCP, or explore the TypeScript SDK preview. Ontology validation, consolidation, and policy run inside the managed service for every submitted observation.

Explore the integrations

Common questions

Principles, answered.

Didn’t find your answer?Ask the team Read the documentation
Why are writes treated as proposals rather than edits?

Information can be late, duplicated, incomplete, or conflicting. Consolidation checks each proposal against the model, current state, and policy. The caller sees what was applied, skipped, or held, along with violations, and can act on that result.

Why keep current state and history apart?

Current-state queries read accepted facts directly. Earlier values stay available separately, with the periods when they were recorded as valid. This lets an agent answer a current question and investigate an earlier value without mixing the two.

Does the ontology reject every non-conforming observation?

Open classes accept undeclared fields and report type, allowed-value, and relationship mismatches without blocking unless a field is enforced. Strict classes refuse those writes. An enforced field can reject an invalid type or value on an open class. Missing required fields are advisory in both modes.

How are changes to the schema governed?

Schema changes go through a proposal service that validates the complete candidate schema before applying it. An expected revision can guard against a stale change, and applied changes are audited. Templates are adopted explicitly. Agent permission to change schema depends on the instance’s governance settings; human administration requires authenticated user access.

Why is provenance mandatory?

Source, actor, confidence, timestamp, and extraction method make each mutation inspectable and support conflict decisions and review. They document the evidence behind a recorded fact. The quality of that fact still depends on the observations and policies you provide.

What does it mean that projections are views?

Markdown pages, summaries, and exports are generated from the database. Editing one does not mutate stored state. A correction goes back through the proposal path, where the same ontology, policy, and provenance requirements apply.

Where do these rules run?

Inside the managed ConceptMem service. Clients submit observations and schema proposals through hosted REST or MCP; the TypeScript SDK is in preview. The service validates and applies proposals according to the caller’s authority and workspace rules.

Keep exploring

Explore how the record works.

Start with one shared record

Give your next agent
the current record.

Create a database on the free plan, connect an agent over MCP or REST, and propose your first update.