Home How it works Knowledge Sources
Source control for memory

Bring the knowledge your team
already trusts into Tenure.

AGENTS.md, Confluence, ADRs, READMEs, coding standards, and manual beliefs can all feed into memory. The docs stay where they are. The beliefs are what your AI uses.

Docs as sources

Keep your documentation.

Tenure does not replace AGENTS.md, ADRs, or Confluence. It turns stable parts of them into scoped beliefs with provenance.

Manual control

Create beliefs directly.

Teams can add explicit policies, conventions, decisions, preferences, and constraints without waiting for the system to infer them.

Safe adoption

Import does not mean inject.

Start in document-driven, observation, or curated modes so teams can review memory before models act on it.

Supported source types

Tenure starts with the knowledge you already have.

Most teams already have what a model would need to know. It's just scattered across files, docs, and old decisions that no tool can pull from consistently.

#

AGENTS.md and CLAUDE.md

Use existing agent instructions as trusted source material, then convert stable guidance into scoped runtime beliefs.

Confluence and internal docs

Extract conventions, architecture rules, and team knowledge from documentation your organization already trusts.

ADR

ADRs and decision logs

Turn architectural decisions into model-facing state so future sessions inherit the why, not just the current code.

{ }

README and repo docs

Seed project-specific facts, setup rules, service boundaries, and codebase conventions.

Coding standards

Make lint preferences, testing rules, review expectations, and generated-code conventions available per turn.

+

Manual beliefs

Create explicit beliefs for policies, preferences, decisions, and constraints that should not depend on inference.

Markdown is a source

AGENTS.md is useful on its own. It just can't carry your whole memory system.

Markdown files are great for instructions a person can read. Tenure adds the scope, source tracking, review step, and per-turn control that markdown alone can't give you.

Markdown-only memory

Everything sits in one file.

  • Hard to know what was actually used.
  • No per-turn audit trail.
  • Context grows stale unless manually curated.
  • Different tools may read different files.
Tenure memory

Docs become governed state.

  • Source docs remain linked as provenance.
  • Beliefs are scoped by user, team, project, or org.
  • Only relevant beliefs are injected per turn.
  • Teams can review, edit, pin, retire, or reject beliefs.
Manual belief creation

Some things shouldn't have to wait to become memory.

If a rule is already known, your team shouldn't need repeated conversations to teach it to the model. Admins, team leads, and developers can just add it directly.

New belief
typedecision
canonical_namebilling_service_owns_invoice_state
contentThe billing service is the source of truth for invoice state.
why_it_mattersModels should not add invoice state writes to downstream services.
scopeteam:payments · project:billing-api
sourceADR-042 · created by human
Good manual beliefs
  • Architecture boundaries the model must respect.
  • Security policies that should override user preference.
  • Team conventions that are already settled.
  • User preferences that should persist across tools.
Better as docs only
  • Long procedural guides that should stay human-readable.
  • Reference material that changes constantly.
  • One-off troubleshooting notes.
  • Information the model should search, not assume.
Safe rollout

Start with what you trust. Add more once it proves out.

Trying to let the AI learn everything at once isn't the fastest path. Seeding from sources you already trust, watching what Tenure pulls out, then approving what's worth keeping gets you there faster.

01

Import sources

Start with AGENTS.md, ADRs, coding standards, and Confluence pages that represent stable team knowledge.

02

Extract candidates

Tenure proposes structured beliefs with type, scope, confidence, source, and why-it-matters fields.

03

Review and approve

Pin, edit, reject, or approve before those beliefs become model-facing memory.

04

Inject per turn

Only relevant approved beliefs are injected into the model, with audit trails showing what was used.

In practice, this means keeping the docs your team already trusts and using Tenure to make the durable parts of them inspectable, scoped, and available across every AI client.

The point

Knowledge sources are where memory starts.
Beliefs are what the model uses.

Tenure lets your team decide what feeds into memory, keep track of where it came from, and control what actually reaches the model.