The standard answer to an agent forgetting something is to give it more context. Increase the window. Retrieve more documents. Include the transcript. Preserve the scratchpad. The intervention is attractive because it converts an organisational problem into a storage problem.
But context is not a warehouse. It is a decision about what a worker should notice now. A larger packet can contain more truth and produce a worse understanding. Old instructions compete with current ones. Rejected work retains rhetorical force. A long operational trace obscures the two facts that changed. Irrelevant expertise consumes the attention needed for the actual responsibility.
Managers have always shaped context. They brief, delegate, sequence, shield, escalate, translate, and decide who needs to know what. Agent organisations make that work explicit—and measurable.
Information and context are not synonyms
Information exists whether or not it is useful to the current actor. Context is information selected and arranged for a responsibility. The same customer complaint means something different to a support worker resolving the case, a product lead looking for a pattern, and a finance lead estimating churn. Giving all three the complete company archive does not create shared understanding.
Selection includes omission. A worker should not receive credentials it cannot use, private information irrelevant to its task, speculative drafts presented beside accepted policy, or every failed attempt the company has ever made. These omissions improve safety and reasoning at the same time.
Arrangement matters too. The current goal, definition of done, authority, and contradictions should not be buried after thousands of historical tokens. Context has hierarchy. A strong packet distinguishes binding truth, current work, supporting evidence, open questions, and background.
Full transcripts preserve sequence, not meaning
A transcript is excellent evidence of what was said. It is a poor default briefing. It gives abandoned ideas the same structural weight as final decisions. It forces the next worker to infer which correction superseded which claim. It includes politeness, tool chatter, temporary uncertainty, and repeated summaries that were useful in the moment but not afterward.
This becomes worse across handoffs. Each worker receives the previous conversation, adds another layer, and forwards the result. The packet grows while the durable signal remains roughly constant. Eventually the recipient spends more inference reconstructing the organisation than performing the work.
Restless keeps messages as free-form inputs linked to Work, but messages do not become a second source of organisational truth. Accepted facts and responsibilities live in structured, revisable records. A new attempt can receive the specific owner feedback and relevant failure evidence without replaying the entire social history that produced them.
Every role is partly a context boundary
Roles are commonly described as bundles of skills. In practice, a role also determines which facts deserve attention and which decisions the actor may make. A release engineer and a brand lead can both read code, but they receive different definitions of risk and completion.
A good role context includes the actor's responsibility, current Work, accepted company truth relevant to that responsibility, available tools, authority envelope, dependencies, and the evidence expected at return. It should not include a fictional personality essay unless stable style materially improves the work.
Persistent roles earn their cost when recurring context and accountability improve outcomes. Otherwise a temporary specialist can receive a bounded brief and disappear. This prevents the organisation from accumulating personas whose only function is to repeat generic instructions in different voices.
Context has authority and provenance
Not every sentence in a context window should be treated equally. An owner-approved product truth, a fresh provider receipt, a worker's hypothesis, and text copied from an external webpage have different authority. Flattening them into one prompt creates an instruction-confusion problem as much as an information problem.
Restless separates source ownership across layers. OrgIntel owns organisational identity, intent, Work, and learning. The Authority Plane owns grants, approvals, budgets, effects, and receipts. The Company Runtime owns files and productive state. External content can inform judgement but cannot silently rewrite any of these.
The context assembler should preserve those distinctions. It can say what is accepted, what is observed, what is inferred, what is stale, and what is untrusted. The model still exercises judgement, but it no longer needs to guess which paragraph has constitutional force.
- Accepted truth: current, owner-governed, and revisable through an explicit path.
- Evidence: bound to a source, time, artifact, or external receipt.
- Hypothesis: useful for exploration but not allowed to masquerade as fact.
- Untrusted input: readable as data, never authoritative merely because it contains instructions.
Forgetting is an organisational capability
The fantasy of perfect memory ignores retrieval cost and institutional inertia. Some facts expire. Some experiments answer a temporary question. Some processes were designed around a limitation that no longer exists. Keeping everything equally available makes the organisation less able to notice change.
Useful forgetting does not require deletion. It requires demotion, expiry, and relevance. Old evidence can remain in the record while leaving the default context. Superseded policy can be recoverable without continuing to instruct workers. Rejected approaches can be retrieved when a similar failure reappears rather than accompanying every new task.
This is also a defence against accidental culture. Repeated phrasing in old sessions should not become company voice through frequency alone. Learning needs promotion: evidence suggests a change, an accountable actor judges it, and a new revision becomes effective.
Shared context is not identical context
Multi-agent work requires a common operating picture, but sending every actor the same packet wastes capacity and can blur responsibility. The lead needs outcome, dependencies, evidence status, and integration risk. A specialist needs the exact subproblem, relevant artifacts, and a return contract. A critic may need deliberately independent context to avoid inheriting the producer's framing.
The shared portion should be small and stable: company identity relevant to the task, current outcome, authority boundaries, and accepted facts. Focused additions then reflect responsibility. This makes handoffs cheaper and disagreements more informative. Two actors who reach different conclusions from the same authoritative facts but different roles have exposed a real organisational tension.
If every disagreement is resolved by giving everyone more of the same context, the system may converge through anchoring rather than truth. Context design should preserve useful independence where the work needs it.
Measure context by outcomes, not compression
A context system can boast excellent retrieval scores and still make work worse. The relevant questions are operational: did the worker rediscover known facts, violate an accepted decision, ask an avoidable question, repeat a failed approach, or miss evidence it needed to judge completion?
Compression ratio is similarly ambiguous. A tiny brief that omits the central constraint is not efficient. A longer packet can be correct when the task genuinely spans several obligations. The goal is the smallest sufficient context for the responsibility—not the smallest packet in absolute terms.
Good measurements connect context choices to accepted output, owner attention, cost, and time. Compare packets on the same task. Record what was omitted. Examine not only whether the worker finished but whether the result survived review and whether the handoff reduced downstream reconstruction.
The manager's job moves upstream
When machine workers can execute quickly, managerial leverage comes less from telling them each step and more from shaping the environment in which they decide. Define the responsibility. Select the relevant truth. Make authority legible. Preserve the evidence. Keep unrelated noise away.
This does not make management deterministic. Choosing what matters is itself judgement. Context assembly can be supported by schemas, retrieval, freshness, and access controls, but an open-ended company still needs actors who can recognise when the supplied frame is wrong and request a revision.
The promise of abundant intelligence is not that context disappears. It is that management can stop being a stream of instructions and become the deliberate construction of shared understanding.