The Promotion Pipeline: From Observation to Trusted Knowledge

The single most important rule for safe agent memory is also the simplest to state: do not write every observation directly into long-term, trusted memory. What an agent just saw is raw, unverified, and possibly planted. Treating it as durable truth is how a store quietly fills with confident errors. Instead, run observations through a promotion pipeline, where an item earns its way from a raw note to trusted knowledge by clearing a check at each stage, and is retired when it no longer holds. The pipeline is what turns remembering into knowing.

Don't write straight to trusted

The tempting shortcut is to save whatever the agent observed as a fact and reuse it later. It feels efficient and it is quietly dangerous, because a single unverified observation, a misread result, or a poisoned input becomes a durable belief the agent will act on.

The discipline is to separate the act of noticing something from the act of trusting it. Noticing is cheap; trusting should be earned.

The promotion pipeline

A safe pipeline has clear stages. A raw observation is captured. It becomes an attributed fact once you record its source, time, and scope. It becomes validated memory when a deterministic check, a trusted source, or a human confirms it. Only then is it promoted to knowledge available for future retrieval. And when it goes stale or is contradicted, it is decayed or retired.

Each stage is a gate. Nothing skips ahead, so nothing reaches trusted knowledge without having been attributed and validated first.

An item moves from raw observation to trusted knowledge only by clearing each stage.

Each stage earns the next

Attribution is the cheapest and most valuable step: recording where a claim came from, when, and how confident you were. It turns an anonymous assertion into something you can later judge or revoke.

Validation is the higher bar: a test passed, a trusted source agreed, a human approved. Promotion should require it, so that trusted knowledge is by definition knowledge something other than the model has confirmed.

Most observations never promote

A healthy pipeline is a funnel. Many observations are captured, fewer are worth attributing, fewer still survive validation, and only a small fraction become trusted knowledge. That narrowing is the point, not a inefficiency.

If almost everything an agent notices ends up as trusted memory, the pipeline is not filtering; it is just relabeling raw observations as facts, which is exactly the failure it exists to prevent.

A healthy pipeline promotes only a small fraction of what the agent observes.

Promote only what is observed and validated

Promotion sits at an intersection. Something the agent merely observed is not enough; something validated in the abstract but not actually seen in this system is not enough either. Only where an item is both observed and independently validated should it become trusted knowledge.

Holding to that intersection keeps the trusted store small, accurate, and defensible, rather than bloated with plausible-sounding but unconfirmed claims.

Promote only what is both observed in this system and independently validated.

Decay and retire

Promotion is not forever. Trusted knowledge that goes stale, is contradicted, or is superseded must be downgraded or removed, or the store slowly fills with confident falsehoods. So the pipeline runs in both directions: items climb by earning trust and fall when they lose it.

Building decay in from the start is what keeps a memory system honest over months of use, rather than accumulating a sediment of outdated beliefs.

⚠️

Trust is revocable. A promoted fact must be able to fall back down the pipeline when it goes stale or is contradicted.

What a practitioner decides

Designing a promotion pipeline means deciding the stages, the check that gates each one, and the rules for decay. What must be attributed, what counts as validation, what may be promoted, and when trust is revoked are all explicit choices, not defaults.

The question for anything an agent wants to remember as true is not did we see it. It is: has it been attributed, validated, and promoted, and is it still current?

Key takeaways

Never write a raw observation directly into trusted, long-term memory.
Separate noticing from trusting: run observations through a promotion pipeline.
Stages are gates: raw, attributed, validated, promoted, then decayed or retired.
A healthy pipeline is a funnel; only a small fraction of observations should be promoted.
Promote only what is both observed here and independently validated.
Build in decay so trusted knowledge falls back down when it goes stale or is contradicted.

Glossary

Promotion pipeline: A staged process that moves a memory item from raw observation to trusted knowledge by clearing checks.
Raw observation: Something the agent just noticed, unverified and possibly planted.
Attribution: Recording a claim's source, time, scope, and confidence.
Validation: Confirming a claim via a deterministic check, a trusted source, or a human.
Promotion: Making a validated item available as trusted knowledge for future retrieval.
Decay: Downgrading or removing a memory when it goes stale, is contradicted, or is superseded.
Trusted knowledge: Memory confirmed by something other than the model, safe to act on.