This is Part 2 of Memory Is Not a Database. Part 1 develops the evidence-first memory model; this part asks what happens when memory, users, sources, tasks, and agents all grow—but not at the same rate.
TL;DR: “Scale” is not one number. Memory volume, users, sources, concurrent tasks, agents per user, and agents per task stress different boundaries. Personal memory, task state, agent scratch space, the action journal, and procedural memory need different ownership and lifecycles. Partition by person, coordinate by task, and give each agent the smallest context and authority its role requires.
Scale has multiple dimensions
A personal-agent platform can grow along several largely independent axes:
| Dimension | What grows | Primary pressure | Architectural response |
|---|---|---|---|
| Memory per person | Years of conversations, events, documents, actions, and derived beliefs | Retrieval precision, consolidation, staleness, and storage tiers | Keep a compact current model; incrementally summarize and consolidate; retain cold evidence for audit and re-derivation |
| Number of people | Independent memory spaces and workloads | Isolation, throughput, noisy neighbors, compliance, and regional placement | Partition by subject or household; enforce tenant boundaries; apply per-user quotas and regional affinity |
| Sources per person | Gmail, Calendar, Docs, browsers, commerce, finance, devices, and future providers | Identity resolution, duplicates, conflicting claims, permissions, and deletion | Normalize into a common evidence envelope while preserving source authority and lineage |
| Tasks per person | Concurrent reminders, research, purchases, plans, and long-running goals | Scheduling, priority, cancellation, contention, and context switching | Give every task durable state, an owner, a budget, a deadline, and explicit dependencies |
| Agents per person | Multiple specialist agents acting for the same person | Inconsistent beliefs, duplicate work, write conflicts, and excessive context fan-out | Share one governed personal memory while giving each agent scoped access and private working state |
| Agents per task | A coordinator delegating research, browsing, planning, verification, and execution | Coordination overhead, partial failure, result merging, and runaway cost | Use a task graph, leases, idempotent steps, bounded fan-out, and a shared task workspace |
| Actions and observations | Tool calls, browser events, retries, intermediate outputs, and outcomes | Event volume, auditability, and memory pollution | Keep the full action trace in the audit plane; admit only useful evidence and promote only validated procedures |
| Geography and availability | Regions, devices, intermittent connections, and uptime expectations | Latency, residency, failover, and consistency | Place hot state near the person, replicate deliberately, and distinguish local responsiveness from authoritative commits |
| Model and reasoning demand | Extraction, reflection, reranking, planning, and agent coordination | Cost, latency, rate limits, and nondeterminism | Batch background inference, cache compiled context, use smaller models where sufficient, and allocate explicit reasoning budgets |
Scaling one axis does not imply scaling the others. Ten million lightly active users create a different system from ten thousand users with a decade of history, or one power user running fifty concurrent agents. Capacity planning therefore needs a workload vector, not a single “users” number.
Memory splits into scopes
At small scale, everything can look like one assistant history. At larger scale, five scopes become distinct:
- Personal memory is long-lived knowledge about the person. It belongs to the user, not to any individual agent.
- Task state records a goal, plan, artifacts, decisions, dependencies, and status. It is shared only by agents working on that task.
- Agent working state is temporary scratch space for one agent or subtask. It usually expires after the result is merged.
- The action journal is the authoritative audit of what every agent attempted and what happened.
- Procedural memory contains reusable strategies promoted from successful trajectories. It does not inherit every task trace automatically.
Confusing these scopes causes predictable failures: an agent’s guess becomes a user belief, a failed task pollutes future context, every specialist receives sensitive history, or a one-off workaround becomes permanent procedure.
One user can have many tasks and agents
The natural horizontal partition remains the person or child because different people can usually process independently. Within that partition, work is scheduled by task and coordinated as a dependency graph.
When one task needs multiple agents, a coordinator compiles a minimal task packet for each specialist rather than copying the person’s complete context to all of them. Specialists receive only the memories, source access, tools, budget, and authority their roles require. They publish typed results into a shared task workspace and append actions to the audit journal.
The coordinator resolves conflicts, validates completion, and records the outcome as evidence. Durable personal beliefs or procedures are promoted later through the governed memory pipeline—not written directly by whichever sub-agent finishes first.
Agent fan-out must earn its coordination cost. Parallel independent research, specialized tools, or adversarial verification can justify multiple agents. A serial task with heavily shared context often becomes slower and more expensive when split.
Ordering and consistency are local decisions
Ordering need not be global. Independent searches and document analyses can run concurrently. Identity changes, permissions, financial actions, user corrections, and belief revisions may require strong sequencing within a subject or task.
Eventual consistency is acceptable for newly inferred understanding. It is not acceptable when applying revoked consent, deleting source-derived knowledge, enforcing approvals, transferring task ownership, or committing consequential external actions. Those boundaries should be explicit rather than inherited accidentally from the storage system.
Four operating rules follow:
- Keep synthesis off the live path: admit evidence and retrieve prepared context synchronously; derive richer understanding asynchronously.
- Keep views re-derivable from the journal under recorded model, prompt, and policy versions.
- Budget and isolate work per user and per task so one person’s agent swarm cannot starve another’s.
- Cache compiled context by subject, task, audience, policy, and memory version—not merely by query text.
Geography, cost, and evaluation
Hot personal state and processing should usually live near the person, while residency and deletion obligations constrain replication. Large encrypted artifacts can move to colder object storage; the current user model and active task state remain hot. Regional failure may permit a degraded read path, but authoritative actions still require a clear commit boundary.
At scale, raw storage is unlikely to be the hardest problem. The costly and risky work is deciding what deserves admission, reconciling evidence across sources and agents, retrieving the right context for each subtask, and determining whether added memory or added agents improve outcomes enough to justify their model and coordination cost.
The evaluation suite therefore matters as much as the storage engine. It should measure correct recall, false beliefs, stale-memory use, contradiction resolution, procedural transfer, context precision, cross-user isolation, latency, cost, task success, and the marginal value of agent fan-out.
The target is not infinite memory or infinite agents. It is bounded cognition: the right evidence, understanding, authority, and collaborators for the task at hand.