Run Memory¶
Run memory is ORCA Framework's durable record of project-specific facts, decisions, and context that should help future runs act consistently.
Purpose¶
Use run memory to preserve:
- stable product or repo decisions
- recurring user preferences
- known constraints and blockers
- operational context that future runs should not rediscover
- durable legacy-system facts, business rules, and modernization constraints
- goal lifecycle state, objective, blocker, and steering notes when useful for resume
What Run Memory Is Not¶
Run memory is not a full execution log.
If you need to know what happened during a specific run, use a trace. See docs/observability.md.
If you need the current coordination picture for an active run, use shared state. See docs/shared-state.md.
Good Run Memory Entries¶
Good entries are:
- durable enough to matter later
- short enough to re-read quickly
- explicit about whether they are facts, decisions, or assumptions
Compactness matters. Verbose run-memory entries increase future prompt mass and weaken the point of having a reusable durable memory layer.
Contract¶
Use templates/contracts/run-memory-contract.md when designing memory artifacts or checking whether a memory entry is complete enough for future reuse.
For portable machine-readable exchange, use the run-memory schema in schema/versions/v1/run-memory.schema.json.
Relationship To Workflow¶
- onboarding discovers context
- legacy archaeology extracts behavior, business logic, and modernization risk
- spec and plan define intended work
- goal contracts track durable objective execution when used
- run memory preserves durable lessons
- traces record what happened during a run
- receipts summarize the run outcome for quick reinspection
- lineage links the artifact to upstream and downstream workflow state
- shared state records the current multi-role picture for an active run
- controller and executor role decisions may be preserved when they become durable operating context
- evals judge how well the workflow performed
- workflow metrics record time, retries, and optional usage signals
- token-efficiency policy should bias memory toward compact summaries rather than transcript replay
Run memory may reference benchmark results, workflow metrics summaries, regression-task decisions, legacy audit findings, goal state, or checkpoint outcomes when those become durable project context, but it should not duplicate the underlying per-run artifacts.
Portable schemas extend run memory by making the required shape explicit for validation or mapping. They do not replace concise human-readable entries.