Workspace memory example for run continuity and handoffs

[SuhaibAslam]

Hi Symphony maintainers,

I wanted to share a small example proposal around workspace-local memory for run continuity and
handoffs.

Symphony already preserves per-issue workspaces across runs, retries, and continuation. One gap I
noticed is that there is not yet an example of a compact, structured state artifact that future
runs or human reviewers can use to recover the useful reasoning state from prior attempts.

The practical problem this tries to address:

• retries can rediscover decisions already made
• continuation runs can lose unresolved questions or known risks
• human reviewers may need to reconstruct context from raw logs
• implementation handoffs can lose validation state and failed-attempt notes

I put together a small examples-only branch here:

https://github.com/SuhaibAslam/symphony/tree/codex-workspace-memory-example/examples/workspace-memory

It adds:

• examples/workspace-memory/README.md
• examples/workspace-memory/WORKFLOW.md
• examples/workspace-memory/WORKSPACE_MEMORY.json

The example is intentionally non-normative. It does not propose a database, a vector store, or native
orchestrator writes. It just demonstrates a workspace-local JSON artifact with fields like:

• summary
• decisions
• rejected directions
• open questions
• known risks
• failed attempts
• validation
• next actions
• handoff notes

The idea is that a future retry or continuation run could load this compact state rather than
reconstructing prior context from raw logs or hidden model history.

I would be curious whether this kind of examples-only contribution would be useful for the repo, or
whether you would prefer this direction to stay out-of-tree unless/until there is native
implementation support.

If it is useful, a possible follow-up could be an optional SPEC.md extension that defines path
safety, size limits, prompt exposure, and observability semantics. But I think the example is the
right first step because it keeps the proposal concrete without standardizing too early.

[SuhaibAslam]

I turned this idea into a more complete runnable branch here:

https://github.com/SuhaibAslam/symphony/tree/codex-cognitive-harness

The framing moved from workspace memory alone to a broader Cognitive Harness pattern: state, rationale, reflection, review, and continuation artifacts inside a Symphony-style workspace.