Memory Without Silent Authority
Everything Claude Code (ECC) treats continuity as a problem of bounded context, not as permission to preserve every conversation. This chapter focuses on two different forms of persistence: the Memory Vault, which carries facts, handoffs, and notes across sessions and harnesses, and continuous learning, which extracts repeated behavior into scoped instincts. The practical judgment is how to preserve what the next agent needs without allowing remembered text or learned preferences to become invisible policy.
A coding task rarely ends exactly when the terminal closes. A migration may be half-tested, a review may have identified a risk that has not been fixed, or a decision may depend on an external work item. The next session needs continuity, but a raw transcript is a poor substitute for a current state. It contains abandoned approaches, duplicated context, tool noise, and instructions that were relevant only at the time they were written. More importantly, it is difficult to tell which sentence is a verified fact, which is a guess, and which is merely an intention.
A bounded handoff preserves the next decision
ECC's Memory Vault uses human-readable ecc.memory.v1 Markdown documents as a shared context layer. Project and team memories live with the repository's local memory surface; user memories live in a separate user scope. The same documents can be reached through the ecc memory CLI or an optional local stdio MCP server, allowing Claude Code, Codex, OpenCode, Cursor, and other harnesses to exchange a portable handoff without copying vendor-specific transcripts.
The important design choice is not the file format alone. It is the trust ceiling attached to the format. A new memory is created as unreviewed and is never silently promoted into a rule, skill, policy, or architectural decision. Human acceptance does not rewrite the memory's trust field. Instead, verified knowledge is promoted into a governed project artifact such as documentation, a decision record, a runbook, or a rule. The Vault can point toward authority; it does not impersonate authority.
Consider a session that has completed the first part of an authentication migration. A useful handoff might state the objective, the tests already run, the files or work item involved, the remaining blocker, and the next concrete action. It can say that token-rotation tests pass while rollout is still pending. That is more useful than preserving every search result, and safer than writing “authentication is ready” when the evidence supports only one slice of the migration. The next harness can search for the handoff, read it directly, and verify its important claims against the repository or test output.
The Vault's scopes make that choice explicit. Project memory is appropriate for repository-local context. Team memory can support human-inspected sharing, but committing a file to Git does not make its contents authoritative. User memory follows the operator across repositories and is recalled only when explicitly requested.