Update, September 18, 2026. Skyflo 1.1.0 changed the rule this essay argues for. Memory the evidence supports (your own words, or a command the work ran successfully) is now kept without a click, and only weaker records wait for review. Correcting a suggestion before keeping it ships, and "Dreamer" did not; consolidation now runs after each completed mission instead. Skyflo Improves How It Improves explains why. The essay below is kept as written.
The Tax Nobody Itemizes
Here is a cost that appears on no dashboard: every time an engineer or an agent starts work on your system, they pay a context tax. Which repositories matter. Who owns the keys. Why the build is pinned to that base image. What broke the last time someone touched the worker rollout. None of it is in the ticket. All of it is required.
For humans, the tax is paid in interruptions. For AI agents, it is paid in re-explanation. The operator supplies the same background again because a fresh session does not automatically inherit the source and decision that mattered last time.
Skyflo currently supports personal, source-linked memory on the linked Mac. A mission proposes a candidate and the user accepts or dismisses it. Only accepted memory is retrieved later. Organization-wide memory is Planned. Dreamer and inline correction are Preview. The status page is the authority for those boundaries.
The strange part is that engineering organizations do write things down. The problem is where.
Context Has a Half-Life
Engineering knowledge today is stored in four places, and each one leaks:
- Chat threads hold the decision and the debate, then become hard to retrieve.
- Closed pull requests hold what changed and how, while the reason may live in a review comment or nowhere.
- Runbooks and wikis hold how the system worked on the day someone wrote the page. Their accuracy decays from that moment, silently, with no indicator.
- Senior engineers hold the real integration map: ownership, history, and the scar tissue of past incidents. That map can leave with them.
The pattern across all four: knowledge is written where the work happened to occur, not where the next person will look. Retrieval depends on remembering that the answer exists and where it was left. That works at ten engineers. It visibly fails at fifty.
The fix is not "write more documentation." Documentation is a fifth leaky bucket with better formatting. The fix is changing what produces the record and what consumes it.
Memory as a Proposal, Not an Automatic Write
The important product boundary is that durable memory is not written automatically when a mission completes.
A mission can propose source-linked memory candidates as part of its close path. The user decides which candidates become persistent personal memory. A rate-limiting mission might propose records such as:
- Outcome: rate limits enforced at api-gateway via token bucket in Redis; per-tier limits live in
platform-infra/limits.yaml. - Constraint: 429 responses must carry
Retry-After; v2 SDK clients retry automatically, v1 clients surface the error. - Ownership: platform-infra owns the Redis cluster; billing-service owns tier definitions.
When a later mission touches adjacent work, it can retrieve the accepted records with their sources. Dismissed candidates do not become memory, and accepted records can be forgotten or deleted.
Personal accepted memory is Available. Organization-wide memory, including shared architecture and ownership records, is Planned. It should not be described as current team behavior until the product ships the access control, promotion, and governance needed to make shared memory safe.
Verified, Not Vibes
Any long-lived memory system faces the staleness problem: yesterday's fact can become today's bug. A useful record therefore needs provenance and a way to leave circulation.
Two properties keep Skyflo's memory honest:
Records carry provenance. A memory is not a floating sentence; it links to the mission and source material that produced it. An accepted, source-linked record is a different object from an unsupported note.
The user controls retention. Accepted records can be forgotten or deleted. Inline correction is Preview, so the current interface should not be described as supporting edit-then-accept until that control is available.
Dreamer is Preview. Its intended role is to review completed work while the machine is idle and propose consolidated records with sources. It does not make organization-wide knowledge current, and nothing it proposes should become persistent without user acceptance.
Where Memory Lives
Personal memory follows the same rule as the Desktop harness: local execution and storage, with model-provider traffic stated separately.
Personal memory is stored on the linked Mac. Model requests use a supported signed-in harness or go to Moonshot or Z.ai when the user configures that provider's key.
Shared organizational memory is Planned. The moment memory is shared, it becomes infrastructure that needs access control, promotion review, and an audit trail. The pricing catalogue labels that capability rather than presenting it as available.
The Compounding Argument
The strategic case for engineering memory is that it changes the slope, not just the level.
Without memory, every mission pays the full context tax. With accepted, source-linked personal memory, a later mission can start with a relevant constraint and inspect where it came from.
Ask any staff engineer what makes them effective and very little of the answer is typing speed. It is that they remember why things are the way they are, what was tried, and what broke. Engineering memory should make that context retrievable without pretending that every proposed record is already true or shared across the organization.