Memory and State Management
Procedural Memory — Stack
Procedural memory › Stack
Technology landscape as of 2026-10-10. Knowledge cutoff May 2026.
Selection criteria
You are choosing how re-usable know-how is represented, how safely it can be replayed, and how quickly a bad one can be taken out of circulation.
- Replay safety. Whether the representation can be executed without a human reading it first, and what the containment is when it does the wrong thing.
- Parameterisation. Whether the mechanism separates the shape of a routine from the values it was captured with, since a routine that carries last week's record identifiers is a hazard.
- Versioning and rollback. Whether a routine has a version that can be pinned, compared, and withdrawn, or whether editing it changes behaviour everywhere immediately.
- Outcome attribution. Whether the system can tell which routine was used on a run, which is the precondition for ever demoting one.
- Environment binding. Whether a routine can be tied to the deployment, tenant, or version it was learned against, and refuses to run elsewhere.
- Review surface. Whether a human can read, diff, and approve a routine before it becomes something the agent will apply unsupervised.
- Invalidation latency. How long it takes for a withdrawal to take effect across every agent that may have cached it.
Options
| Option | Best for | Watch out for |
|---|---|---|
| Prose playbooks retrieved into the prompt | Open-ended tasks where the environment changes faster than any script could | Advisory only; the model may ignore or misapply it, and there is no way to tell which happened |
| Versioned code or scripts in a repository, retrieved by task match (generated functions, Voyager-style skill libraries) | Deterministic, repeatable operations against a stable interface | Generated code is an execution surface; sandboxing, review, and dependency drift all become your problem |
| Parameterised call sequences in a declarative format (a JSON or YAML step list the runtime executes) | Routines over a known tool set where replay must be auditable | Expressiveness runs out quickly; conditionals and error handling push it towards being a programming language |
| Workflow definitions in an orchestration engine (Temporal, Airflow, Step Functions, n8n) | Routines that are long-running, need retries, and must survive process restarts | Heavyweight for small lessons; authoring is a deploy, which defeats the point of learning one at runtime |
| Few-shot exemplars of successful trajectories retrieved by similarity | Improving model behaviour without committing to executable artefacts | Exemplars teach style as much as substance; a bad one is hard to attribute and harder to withdraw |
| Agent-framework skill or tool registries that accept runtime registration | Agents that genuinely compose new capabilities while running | Runtime registration means the tool surface differs between replicas; naming collisions and stale registrations are common |
| Managed prompt and routine registries with versioning (prompt-management platforms, feature-flag systems used for the same purpose) | Teams that need review, staged rollout, and instant withdrawal of a routine | Another control plane in the hot path; a withdrawal is only as fast as its propagation and cache settings |