Floris Vossebeld

The memory promotion pipeline

4 min read

Raw experience shouldn’t become shared knowledge automatically. It gets promoted through stages, shedding specificity and gaining trust at each one. Each arrow below is a gate: knowledge clears it before it earns a wider audience, or it stays put.

The memory promotion pipeline Five stages drawn as a descending funnel: raw episode, private memory, sanitized lesson, approved playbook, and reusable skill. Each bar is shorter than the one above it because every gate rejects most of what reaches it. Between the stages sit four labelled gates: scope to one tenant, strip identifying detail, human review and provenance, and compress to a default. Raw episode private detail scope to one tenant Private memory user or tenant strip identifying detail Sanitized lesson generalised human review + provenance Approved playbook reviewed, signed off compress to a default Reusable skill shareable
The funnel narrows because each gate rejects most of what reaches it, not because the bars count anything. Every step strips specificity and adds review; few episodes reach the bottom.

The climb is deliberate. A raw episode is what actually happened this session, private detail and all, and it stays local. Promote it and it becomes a private memory, a durable note still scoped to one user or tenant. Strip the identifying detail and you have a sanitized lesson, generalisable and carrying nobody’s name. Once a human signs that off with provenance attached, it is an approved playbook. Compress the playbook into a procedure the agent reaches for by default, and it is a reusable skill.

Most episodes never leave the first stage. That’s the point. Promotion is where abstraction and human review happen, which is where leaks and bad lessons get caught. Skip the checkpoints and you’ve rebuilt the leaky shared store from federated-memory-for-enterprise-agents.

What actually does the climbing

The stages are the ladder; they don’t say what moves a lesson up one. For the early, unsupervised steps the answer is increasingly a consolidation pass: a background job that fires after a session goes quiet, reads the raw traces plus whatever’s already in memory, and rewrites the store. Duplicates merged, contradictions resolved, stale entries dropped. Anthropic ships this as a feature literally called Dreams; the Azure SRE agent does the same thing about thirty minutes after a thread goes idle.

The detail I care about is where it stops. Dreams never edits the input store; it produces a candidate the agent’s owner can review and discard. Same boundary as the ladder: the machine can dedupe and generalize on its own, but the step from sanitized lesson to approved playbook is still a human gate. Consolidation automates the climb right up to that line and no further. Treat it as auto-promotion past the gate and you’ve rebuilt the leaky shared store, just with extra steps.

A worked example

Watch one fact climb the ladder:

Raw episode: “Customer X’s integration failed because their internal SAP field Y was misconfigured.” Sanitized lesson: “In enterprise ERP integrations, validate custom field mappings before assuming an API failure.” Reusable skill: “When debugging an ERP integration failure, check authentication first, then field mappings, then validation rules, then downstream workflow triggers.”

Same knowledge, three boundaries. The raw episode names a customer and stays in their tenant. The lesson is true across customers and carries nobody’s name. The skill is a procedure worth running by default, the anecdote gone entirely. Learning without leaking.

Failure modes worth naming

Each gate exists because the pipeline fails in predictable ways:

  • Over-generalisation: a one-off incident promoted into a “rule” that’s wrong most of the time.
  • Hidden leakage: identifying detail surviving sanitization (a customer name buried in an example, a field only one tenant has).
  • Wrong lessons: a confidently-stated playbook that was coincidence, not cause.
  • Stale lessons: knowledge that was true once and quietly expired because nothing retired it.

This is how a workspace’s raw memory turns into durable capability over time (see agent-workspaces), and it’s what makes federated memory more than a set of walls.

Backlinks

Graph View