Home How it works Belief Extraction
Memory lifecycle

Raw chat logs make bad memory.
Tenure pulls out the parts worth keeping.

Conversations, docs, and code context are noisy. Tenure pulls out what's worth keeping, tags where it came from and how confident it is, and gives you a chance to review it before it ever reaches the model.

Structured output

Raw context becomes beliefs.

Tenure extracts decisions, preferences, facts, open questions, expertise signals, and style signals instead of treating entire conversations as memory.

Scoped by design

Memory lands where it belongs.

Project, domain, team, and user scopes keep one task’s context from leaking into unrelated sessions.

Mode aware

Extracted, not always shown.

In observation or curated modes, Tenure can extract and show what it learned without automatically giving it to the model.

The extraction path

Not everything Tenure extracts becomes memory right away.

Tenure keeps capturing signal separate from trusting it. Extraction just creates candidates. What happens after that comes down to validation, scope, your memory mode, and merge policy.

01

Observe source

Tenure watches the relevant context: a conversation turn, onboarding input, imported documentation, or IDE workspace context.

02

Emit sidecar

The model proposes structured memory objects instead of free-form notes: new beliefs, updates, open questions, aliases, and style signals.

03

Validate and scope

Tenure parses, repairs when possible, drops invalid output, and enforces the authoritative project scope in code.

04

Route by mode

Depending on your memory mode, extracted beliefs are ignored, saved as suggestions, merged automatically, or used only for insight.

What gets extracted

Tenure looks for things the model can use later.

A belief isn't just an interesting sentence someone said. It's specific enough that, scoped and injected the right way, it actually changes how the model behaves next time.

Good extraction targets
  • Decisions: “This repo uses Postgres instead of MongoDB.”
  • Preferences: “Prefer small diffs with tests included.”
  • Facts: “Billing events are emitted through the ledger service.”
  • Open questions: “We still need to choose the queue provider.”
  • Expertise: “This user understands MongoDB aggregation deeply.”
  • Style signals: “The team consistently uses explicit return types.”
What should not become memory
  • One-off debugging noise that will not matter later.
  • Pasted reference material treated as if the user believed it.
  • Low-confidence guesses without a durable implication.
  • Project details assigned to the wrong workspace.
  • Temporary instructions that should live only in the current turn.
  • Facts that contradict stronger team or organization policy.
RAG comparison

Extraction is not chunk retrieval.

RAG stores the source text and digs through it again later. Tenure pulls out the actual claim, keeps track of where it came from, and hands the model that one small, useful belief instead of a pile of old transcript.

Chunk memory

Store the conversation.

Later retrieval has to rediscover the useful fact from surrounding noise, stale turns, and unrelated context.

User: We switched validation to Zod...
Assistant: Sure...
User: Actually only at API boundaries...
Belief state

Store the resolved belief.

The model receives the compact, scoped state it can use immediately, with provenance behind it.

{
  type: "decision",
  canonical_name: "zod_api_boundary_validation",
  content: "Use Zod for validation at API boundaries.",
  scope: ["project:billing-api"]
}
Extraction audit
source Conversation turn in billing-api workspace session + turn recorded
candidate “Use Zod for validation at API boundaries.” decision · confidence 0.74
scope project:billing-api enforced by workspace context
mode Curated memory creates a pending suggestion. not injected until approved

The conversion point: Tenure can learn what it would remember before you allow that memory to affect model behavior.

The point

Extraction gives you something you can actually manage.

Once your work turns into structured beliefs, your team can review them, merge them, replace outdated ones, audit the history, and decide exactly when they reach the model.