cross-session-memory

Cross-Session Memory

Overview

contextd's ReasoningBank stores reusable strategies and decisions with confidence scoring. The loop is simple: search before solving, record after solving, and give feedback so confidence stays calibrated.

The loop

1. Search before assuming (task start)

memory_search(project_id, query)

Ask "have I solved something like this before?" before re-deriving an approach. Always search before assuming a problem is novel.

2. Record after solving (task completion)

memory_record(project_id, content, ...)

Capture the why, not just the what. A good memory includes:

• The problem and the approach that worked
• Rejected alternatives and the tradeoff that decided it
• Consequences / gotchas to watch for

3. Report outcomes and feedback

• memory_outcome — after acting on a memory, report whether the task succeeded. This is the reinforcement signal.
• memory_feedback — rate a specific memory as helpful or not, adjusting its confidence.

4. Consolidate (periodically)

memory_consolidate merges similar memories into refined summaries so the bank stays sharp instead of accumulating near-duplicates.

What makes a good memory

Good	Avoid
"Use payload isolation, not filesystem, for multi-tenant vectorstore — avoids N collections; rejected per-tenant DB because of open-file limits."	"Fixed the bug."
Decision + rejected alternative + consequence	Restating code that's already in the diff

When NOT to record

• Information already obvious from the code or docs.
• Secrets or credentials (contextd scrubs responses, but don't author them into memories).