How ReacherX remembers instructions and learns from work

The stable key allows a correction to address the same topic. Scope distinguishes a general workspace instruction from something that applies only to one person.

convex/lib/workspaceMemoryScope.ts resolves context from prospect threads, setup sessions, and workspace threads. Don't replace that with whichever workspace happens to be first in a query. A valid memory saved to the wrong workspace is still a serious bug.

Turn recorded events into possible lessons

Workflows record events such as qualification completion, outreach approval, task failure, a reply, or an edited draft. convex/workflows/memory.ts queues evaluation by workspace and handles stale work.

convex/lib/learningCore.ts contains the distillation functions. These ask for reusable lessons from qualification, enrichment, outreach, or operator information. convex/lib/memoryEvaluatorCore.ts defines relevant event types and the promotion rule.

The current rule is small enough to read directly:

AUTO_PROMOTE_MEMORY_CONFIDENCE is 0.78 in this revision. These are implementation thresholds, not measured probabilities that a lesson is correct. Inspect the evaluator result and supporting event when judging quality.

An event, an evaluation run, a suggestion, and a promoted memory are separate records. Don't count every event as something △ Agent has learned.

Retrieve the right memory with the right authority

convex/lib/workspaceMemoryCore.ts holds the canonical memory model, applicability checks, ranking, and context construction. It distinguishes active, superseded, and disabled memories, and checks relevant scope, surface, and channel.

Operator instructions have priority over learned context. Retrieval is bounded; it does not paste every historical memory into every prompt. The system also maintains indexing state so a failed embedding operation can be retried without losing the canonical instruction.

This separation is useful for debugging. “The instruction exists” and “the relevant operation received it” are two different claims. Trace both persistence and context assembly.

Evaluate behavior, not just storage

A good test saves an instruction, asks for later work in the same workspace, and checks that the applicable context contains it. Add a second workspace and verify the instruction does not cross over. Test a correction, a disabled memory, and an indexing failure too.

Read these for existing coverage:

• convex/workspaceMemory.integration.test.ts saves instructions and builds workspace context.
• convex/workspaceMemoryIndexRetry.integration.test.ts retries failed memory indexing with backoff.
• convex/memoryEvaluationQueue.integration.test.ts keeps memory evaluation queues FIFO per workspace.

The system can use feedback to change later work. That doesn't guarantee steadily improving outcomes, and it doesn't make inferred lessons true. The user-facing memory guide explains how to steer it; Agent observability exposes records that help people check what happened.